remove tuplize

* add a waitlist for graph rewrite

* cleaner

* one context on spec check

.github/workflows/benchmark.yml

@@ -41,6 +41,18 @@ jobs:
       run: |
         echo "CACHEDB=/tmp/pytest-db-ci.db" >> $GITHUB_ENV
         rm -f /tmp/pytest-db-ci*
+    # TODO: remove this step once all old caches are migrated
+    - name: Migrate old huggingface cache (symlinks break onnxruntime 1.24+)
+      run: |
+        cd ~/Library/Caches/tinygrad/downloads/models 2>/dev/null || exit 0
+        for old_dir in models--*; do
+          [ -d "$old_dir" ] || continue
+          repo_id=$(echo "$old_dir" | sed 's/models--//; s/--/\//g')
+          snapshot=$(ls -1 "$old_dir/snapshots" 2>/dev/null | head -1)
+          [ -n "$snapshot" ] || continue
+          mkdir -p "$repo_id"
+          cp -RLn "$old_dir/snapshots/$snapshot/"* "$repo_id/" 2>/dev/null || true
+        done
     - name: Run pytest -nauto
       run: |
         source /tmp/tinygrad_pytest_ci/bin/activate
@@ -513,6 +525,8 @@ jobs:
       run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - name: Run full CIFAR training steps w 6 GPUS
       run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Test full tinyfs load
+      run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 

.github/workflows/test.yml

@@ -1,7 +1,7 @@
 name: Unit Tests
 env:
   # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '15'
+  CACHE_VERSION: '16'
   CAPTURE_PROCESS_REPLAY: 1
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   PYTHONPATH: ${{ github.workspace }}
@@ -29,9 +29,9 @@ jobs:
         deps: testing_unit
         llvm: 'true'
     - name: Speed Test
-      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
     - name: Speed Test (BEAM=2)
-      run: BEAM=2 CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
 
   docs:
     name: Docs
@@ -207,11 +207,11 @@ jobs:
       run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
     - name: Test device flop counts
       run: |
-        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStats.test_simple_matmul
+        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStats.test_simple_matmul
 
   linter:
     name: Linters
@@ -569,7 +569,7 @@ jobs:
       with:
         key: dsp-minimal
         deps: testing_unit
-        pydeps: "onnx==1.18.0 onnxruntime"
+        pydeps: "onnx==1.18.0 onnxruntime ml_dtypes"
         llvm: "true"
     - name: Set up Docker Buildx
       uses: docker/setup-buildx-action@v3
@@ -581,7 +581,7 @@ jobs:
         load: true
         tags: qemu-hexagon:latest
         cache-from: type=gha
-        cache-to: type=gha,mode=min
+        cache-to: ${{ github.event_name != 'pull_request' && 'type=gha,mode=min' || '' }}
     - name: Set MOCKDSP env
       run: printf "MOCKDSP=1" >> $GITHUB_ENV
     - name: Run test_tiny on DSP
@@ -739,7 +739,7 @@ jobs:
           DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
       - name: Run pytest (cuda)
         # skip multitensor because it's slow
-        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore=test/null --ignore test/test_gc.py --ignore test/test_multitensor.py --durations=20
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore=test/null --ignore test/test_multitensor.py --durations=20
       - name: Run TestOps.test_add with PMA
         run: VIZ=-1 PMA=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
       - name: Run process replay tests

examples/mlperf/dataloader.py

@@ -72,7 +72,7 @@ def loader_process(q_in, q_out, X:Tensor, seed):
       #storage_tensor._copyin(img_tensor.numpy())
 
       # faster
-      X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
+      X[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = img.tobytes()
 
       # ideal
       #X[idx].assign(img.tobytes())   # NOTE: this is slow!
@@ -264,8 +264,8 @@ def load_unet3d_data(preprocessed_dataset_dir, seed, queue_in, queue_out, X:Tens
       x = random_brightness_augmentation(x)
       x = gaussian_noise(x)
 
-    X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = x.tobytes()
-    Y[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = y.tobytes()
+    X[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = x.tobytes()
+    Y[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = y.tobytes()
 
     queue_out.put(idx)
   queue_out.put(None)
@@ -379,12 +379,12 @@ def load_retinanet_data(base_dir:Path, val:bool, queue_in:Queue, queue_out:Queue
       clipped_match_idxs = np.clip(match_idxs, 0, None)
       clipped_boxes, clipped_labels = tgt["boxes"][clipped_match_idxs], tgt["labels"][clipped_match_idxs]
 
-      boxes[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = clipped_boxes.tobytes()
-      labels[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = clipped_labels.tobytes()
-      matches[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = match_idxs.tobytes()
-      anchors[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = anchor.tobytes()
+      boxes[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = clipped_boxes.tobytes()
+      labels[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = clipped_labels.tobytes()
+      matches[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = match_idxs.tobytes()
+      anchors[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = anchor.tobytes()
 
-    imgs[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
+    imgs[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = img.tobytes()
 
     queue_out.put(idx)
   queue_out.put(None)

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -2,6 +2,7 @@
 
 export PYTHONPATH="."
 export DEV=${DEV:-AMD}
+export EMULATE="AMD_CDNA4"
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
 
@@ -12,7 +13,7 @@ export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=8 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=1
+export DP=${DP:-8} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
 export GBS=$((BS * GRADIENT_ACC_STEPS))
 
 export MODEL="llama3"
@@ -20,13 +21,14 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="2.5e-4" END_LR="2.5e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
 
-export SEED=5760
+export SEED=${SEED:-5760}
 
-export JITBEAM=3
+export JITBEAM=${JITBEAM:-3}
 export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 
 export FAKEDATA=1 BENCHMARK=10 LLAMA_LAYERS=2

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -13,7 +13,7 @@ export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
+export DP=${DP:-8} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
 export GBS=$((BS * GRADIENT_ACC_STEPS))
 
 export MODEL="llama3"
@@ -21,9 +21,10 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="2.5e-4" END_LR="2.5e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
 
 export SEED=${SEED:-5760}
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/profile.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+export BENCHMARK=5
+export EVAL_BS=0
+export VIZ=${VIZ:--1}
+examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
+PYTHONPATH="." extra/viz/cli.py --profile --device "AMD" --top 20

examples/stable_diffusion.py

@@ -6,7 +6,6 @@ import argparse, time
 from collections import namedtuple
 from typing import Dict, Any
 
-from PIL import Image
 import numpy as np
 from tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit
 from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten, profile_marker
@@ -336,6 +335,7 @@ if __name__ == "__main__":
   print(x.shape)
 
   profile_marker("save image")
+  from PIL import Image
   im = Image.fromarray(x.numpy())
   print(f"saving {args.out}")
   im.save(args.out)

examples/tinychat/tinychat-browser/compile.py

@@ -48,7 +48,7 @@ def prepare_browser_chunks(model):
         weight_metadata = metadata.get(name, default)
         weight_metadata["parts"][part_num] = {"file": i, "file_start_pos": cursor, "size": size}
         metadata[name] = weight_metadata
-        data = bytes(state_dict[name].uop.base.realized.as_buffer())
+        data = bytes(state_dict[name].uop.base.realized.as_memoryview())
         data = data if not offsets else data[offsets[0]:offsets[1]]
         writer.write(data)
         cursor += size

extra/amdpci/am_smi.py

@@ -92,7 +92,7 @@ class SMICtx:
     self.prev_terminal_width = 0
     self.prev_terminal_height = 0
 
-    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:"]
+    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:", "Processing accelerators:"]
     lspci = subprocess.check_output(["lspci"]).decode("utf-8").splitlines()
     self.lspci = {l.split()[0]: l.split(" ", 1)[1] for l in lspci}
     for k,v in self.lspci.items():
@@ -153,7 +153,8 @@ class SMICtx:
     tables = {}
     for dev in self.devs:
       match dev.ip_ver[am.MP1_HWIP]:
-        case (13,0,6)|(13,0,12): table_t = dev.smu.smu_mod.MetricsTableX_t
+        case (13,0,6): table_t = dev.smu.smu_mod.MetricsTableV0_t
+        case (13,0,12): table_t = dev.smu.smu_mod.MetricsTableV2_t
         case _: table_t = dev.smu.smu_mod.SmuMetricsExternal_t
       tables[dev] = dev.smu.read_table(table_t, dev.smu.smu_mod.SMU_TABLE_SMU_METRICS) if dev.pci_state == "D0" else None
     return tables
@@ -279,7 +280,7 @@ class SMICtx:
       device_line = [f"{bold(dev.pcibus)} {trim(self.lspci[dev.pcibus[5:]], col_size - 20)}"] + [pad("", col_size)]
       activity_line = [f"GFX Activity {draw_bar(self.get_gfx_activity(dev, metrics) / 100, activity_line_width)}"] \
                     + [f"MEM Activity {draw_bar(self.get_mem_activity(dev, metrics) / 100, activity_line_width)}"] \
-                    + [f"MEM Usage    {draw_bar((mem_used / mem_total) / 100, activity_line_width, opt_text=mem_fmt)}"] \
+                    + [f"MEM Usage    {draw_bar(mem_used / mem_total, activity_line_width, opt_text=mem_fmt)}"] \
 
       temps_data, temps_data_compact = self.get_temps(dev, metrics), self.get_temps(dev, metrics, compact=True)
       temps_table = ["=== Temps (°C) ==="] + [f"{name:<16}: {color_temp(val)}" for name, val in temps_data.items()]

extra/amdpci/hive_reset.py

@@ -1,12 +1,18 @@
 #!/usr/bin/env python3
 
+import os
 from tinygrad.helpers import Context
 from tinygrad.runtime.support.system import System, PCIDevice, PCIDevImplBase
+from tinygrad.runtime.support.hcq import FileIOInterface
 from tinygrad.runtime.support.am.amdev import AMDev
 
 if __name__ == "__main__":
   gpus = System.pci_scan_bus(0x1002, [(0xffff, [0x74a1, 0x75a0])])
-  pcidevs = [PCIDevice(f"reset:{gpu}", gpu, bars=[0, 2, 5]) for gpu in gpus]
+  for gpu in gpus:
+    drv_path = f"/sys/bus/pci/devices/{gpu}/driver"
+    if FileIOInterface.exists(drv_path) and os.path.basename(os.readlink(drv_path)) == "amdgpu":
+      raise RuntimeError(f"amdgpu is bound to {gpu}. Stopping...")
+  pcidevs = [PCIDevice("AM", gpu, bars=[0, 2, 5]) for gpu in gpus]
   amdevs = []
   with Context(DEBUG=2):
     for pcidev in pcidevs:

extra/amdpci/proclogs.py

@@ -19,8 +19,9 @@ amdev = importlib.import_module("tinygrad.runtime.support.am.amdev")
 amdev.AMDev = AMDFake
 from tinygrad.runtime.ops_amd import PCIIface
 
-def parse_amdgpu_logs(log_content, register_names=None, *, only_xcc0: bool = False):
+def parse_amdgpu_logs(log_content, register_names=None, register_objects=None, *, only_xcc0: bool = False):
   register_map = register_names or {}
+  register_objs = register_objects or {}
 
   def replace_register(match):
     reg = match.group(1)
@@ -37,6 +38,28 @@ def parse_amdgpu_logs(log_content, register_names=None, *, only_xcc0: bool = Fal
   # remove timing prefix
   processed_log = re.sub(r'^\[\s*\d+(?:\.\d+)?\]\s*', '', processed_log, flags=re.MULTILINE)
 
+  # decode register values into field dicts
+  def decode_value(match):
+    reg_name = match.group(1)
+    xcc_part = match.group(2)  # "xcc=0 " or ""
+    val_str = match.group(3)
+    val = int(val_str, 16)
+
+    reg_obj = register_objs.get(reg_name)
+    if reg_obj is not None and reg_obj.fields:
+      fields = reg_obj.decode(val)
+      # show raw for unaccounted bits
+      accounted = 0
+      for name, (start, end) in reg_obj.fields.items():
+        accounted |= (((1 << (end - start + 1)) - 1) << start)
+      unaccounted = val & ~accounted
+      parts = {k: v for k, v in fields.items() if v != 0}
+      if unaccounted: parts['_raw_unaccounted'] = hex(unaccounted)
+      return f"register {reg_name}, {xcc_part}with value {val_str} {parts}"
+    return match.group(0)
+
+  processed_log = re.sub(r'register (reg\w+), ((?:xcc=\d+ )?)with value (0x[0-9a-fA-F]+)', decode_value, processed_log)
+
   # keep only xcc=0 lines (but keep lines with no xcc at all)
   if only_xcc0:
     kept = []
@@ -50,16 +73,18 @@ def main():
   only_xcc0 = bool(getenv("ONLY_XCC0", 0))
 
   reg_names = {}
+  reg_objs = {}
   dev = PCIIface(None, 0)
   for x, y in dev.dev_impl.__dict__.items():
     if isinstance(y, AMRegister):
       for xcc, addr in y.addr.items():
         reg_names[addr] = f"{x}, xcc={xcc}"
+      reg_objs[x] = y
 
   with open(sys.argv[1], 'r') as f:
     log_content = f.read()
 
-  processed_log = parse_amdgpu_logs(log_content, reg_names, only_xcc0=only_xcc0)
+  processed_log = parse_amdgpu_logs(log_content, reg_names, reg_objs, only_xcc0=only_xcc0)
 
   with open(sys.argv[2], 'w') as f:
     f.write(processed_log)

extra/assembly/amd/emu.py

@@ -135,14 +135,6 @@ def _val_to_u32(val: UOp) -> UOp:
   if val.dtype in (dtypes.uint16, dtypes.int16): return val.cast(dtypes.uint32)
   return val.cast(dtypes.uint32)
 
-def _apply_clamp(val: UOp, clmp: int | UOp) -> UOp:
-  """Apply VOP3 clamp modifier: clamp float results to [0.0, 1.0] range."""
-  if isinstance(clmp, int) and clmp == 0: return val
-  if val.dtype not in (dtypes.float32, dtypes.half, dtypes.float64): return val
-  zero, one = UOp.const(val.dtype, 0.0), UOp.const(val.dtype, 1.0)
-  clamped = val.maximum(zero).minimum(one)
-  return clmp.ne(_c(0)).where(clamped, val) if isinstance(clmp, UOp) else clamped
-
 _pcode_fixes = {
   'V_DIV_FMAS_F32': ('D0.f32 = 2.0F ** 32 * fma(S0.f32, S1.f32, S2.f32)',
     'D0.f32 = (exponent(S2.f32) > 127) ? (2.0F ** 64 * fma(S0.f32, S1.f32, S2.f32)) : (2.0F ** -64 * fma(S0.f32, S1.f32, S2.f32))'),
@@ -354,7 +346,7 @@ class _Ctx:
     offset = reg.cast(dtypes.int) * _c(32, dtypes.int) + lane.cast(dtypes.int)
     return buf.index(offset, _lane_active(exec_mask, lane)).store(val.cast(dtypes.uint32))
 
-  def rsrc_dyn(self, off: UOp, lane: UOp | None, bits: int = 32, literal: UOp | None = None, is_f64: bool = False) -> UOp:
+  def rsrc_dyn(self, off: UOp, lane: UOp | None, bits: int = 32, literal: UOp | None = None, is_f64: bool = False, do_cast: bool = True) -> UOp:
     """Read source operand with dynamic offset. Handles SGPR/inline constants (<256), VGPR (>=256).
     If lane is None, only scalar access is supported (off must be < 256).
     is_f64: True for F64 operations where 64-bit literals go in high 32 bits."""
@@ -385,7 +377,7 @@ class _Ctx:
     else:
       scalar_val = sgpr_lo
       if literal is not None: scalar_val = off.eq(_c(255)).where(literal, scalar_val)
-      if bits == 16:  # Float constants: cast F32 to F16
+      if bits == 16 and do_cast:  # Float constants: cast F32 to F16
         scalar_val = is_float_const.where(scalar_val.bitcast(dtypes.float32).cast(dtypes.half).bitcast(dtypes.uint16).cast(dtypes.uint32), scalar_val)
 
     return is_vgpr.where(vgpr_val, scalar_val) if lane is not None else scalar_val
@@ -445,7 +437,7 @@ class _Ctx:
     return UOp.sink(*stores, *self.inc_pc())
 
   def compile_vop_pcode(self, op, srcs: dict[str, UOp], lane: UOp, vdst_reg: UOp, exec_mask: UOp,
-                        opsel_dst_hi: bool | UOp = False, sdst_reg: int | None = None, clmp: int | UOp = 0) -> UOp:
+                        opsel_dst_hi: bool | UOp = False, sdst_reg: int | None = None, clmp: int = 0) -> UOp:
     """Compile VOP instruction. Returns sink with stores and inc_pc."""
     pcode = get_pcode(op)
     vcc_reg = sdst_reg if sdst_reg is not None else VCC_LO.offset
@@ -454,6 +446,24 @@ class _Ctx:
                  'ROUND_MODE': _c(0), 'ROUND_TOWARD_ZERO': _c(0)})  # rounding mode: 0=RNE, RTZ constant
     _, assigns = parse_pcode(pcode, srcs)
 
+    # For integer ops with clamp, compute overflow using wide arithmetic
+    # NOTE: MUL_LO ops don't saturate - they always return the low bits
+    int_saturate = None
+    if clmp and any(p in op.name for p in ('_NC_U', '_MAD_U', '_NC_I', '_MAD_I')):
+      is_signed, is_16bit = '_I' in op.name and '_U' not in op.name, '16' in op.name
+      if not (is_16bit and is_signed):  # Skip 16-bit signed ops due to codegen issues
+        s0, s1, s2 = srcs.get('S0'), srcs.get('S1'), srcs.get('S2')
+        if s0 is not None and s1 is not None:
+          narrow_dt = dtypes.uint16 if is_16bit else (dtypes.int32 if is_signed else dtypes.uint32)
+          wide_dt = dtypes.int32 if is_16bit else dtypes.int64
+          narrow_max, narrow_min = (0xFFFF, 0) if is_16bit else ((0x7FFFFFFF, -0x80000000) if is_signed else (0xFFFFFFFF, 0))
+          def to_wide(x): return (x.bitcast(narrow_dt) if x.dtype.itemsize == narrow_dt.itemsize else x.cast(narrow_dt)).cast(wide_dt)
+          is_sub, is_mad = 'SUB' in op.name, 'MAD' in op.name
+          full = (to_wide(s0) * to_wide(s1) + to_wide(s2)) if is_mad and s2 is not None else \
+                 (to_wide(s1) - to_wide(s0)) if is_sub and 'SUBREV' in op.name else \
+                 (to_wide(s0) - to_wide(s1)) if is_sub else (to_wide(s0) + to_wide(s1))
+          int_saturate = full.clamp(narrow_min, narrow_max).cast(narrow_dt)
+
     raw_stores: list = []
     vcc_val, exec_val = None, None
     for dest, val in assigns:
@@ -468,7 +478,10 @@ class _Ctx:
                        val.cast(dtypes.uint32) if val.dtype in (dtypes.uint16, dtypes.int16) else val.cast(dtypes.uint32) & UOp.const(dtypes.uint32, slice_mask)
             raw_stores.append(('vgpr_slice', (lo_bit, width, val_bits)))
             continue
-        val = _apply_clamp(val, clmp)
+        # For integer ops with clamp, use pre-computed saturated value; for floats, clamp to [0,1]
+        if int_saturate is not None: val = int_saturate
+        elif clmp and val.dtype in (dtypes.float32, dtypes.half, dtypes.float64):
+          val = val.maximum(UOp.const(val.dtype, 0.0)).minimum(UOp.const(val.dtype, 1.0))
         if val.dtype in (dtypes.uint64, dtypes.int64, dtypes.float64):
           lo, hi = _split64(val)
           raw_stores.extend([('vgpr', self.wvgpr_dyn(vdst_reg, lane, lo, exec_mask)), ('vgpr', self.wvgpr_dyn(vdst_reg + _c(1), lane, hi, exec_mask))])
@@ -732,6 +745,7 @@ def _compile_vop3sd(inst: ir3.VOP3SD | ir4.VOP3SD, ctx: _Ctx) -> UOp:
   _, assigns = parse_pcode(pcode, srcs)
 
   has_per_lane_vcc = any('[laneId]' in dest for dest, _ in assigns if dest.startswith('VCC') or dest.startswith('D0.u64'))
+  clmp = getattr(inst, 'clmp', 0)
   if has_per_lane_vcc:
     # VCC computation: RANGE+REDUCE gets axis ID first (lower ID = runs first)
     # This ensures VCC reads source values BEFORE VGPR stores modify them
@@ -743,11 +757,17 @@ def _compile_vop3sd(inst: ir3.VOP3SD | ir4.VOP3SD, ctx: _Ctx) -> UOp:
     final_vcc = ctx.unroll_lanes(get_vcc_bit, exec_mask)
     # VGPR stores: RANGE gets axis ID second (higher ID = runs after VCC loop)
     lane3 = ctx.range()
-    d0_val = None
+    d0_val, vcc_per_lane = None, None
     for dest, val in parse_pcode(pcode, load_srcs(lane3))[1]:
       if dest.startswith('D0') and '[laneId]' not in dest: d0_val = val
+      if dest.startswith('VCC') or (dest.startswith('D0.u64') and '[laneId]' in dest): vcc_per_lane = val
     vgpr_stores = []
     if d0_val is not None:
+      # Apply clamp using carry/borrow bit: ADD overflow->0xFFFFFFFF, SUB underflow->0
+      if clmp and vcc_per_lane is not None:
+        is_sub = 'SUB' in inst.op.name
+        sat_val = _c(0) if is_sub else _c(0xFFFFFFFF)
+        d0_val = vcc_per_lane.cast(dtypes.bool).where(sat_val, d0_val.cast(dtypes.uint32))
       if d0_val.dtype in (dtypes.uint64, dtypes.int64, dtypes.float64):
         lo, hi = _split64(d0_val)
         vgpr_stores.extend([ctx.wvgpr_dyn(vdst_reg, lane3, lo, exec_mask), ctx.wvgpr_dyn(vdst_reg + _c(1), lane3, hi, exec_mask)])
@@ -800,9 +820,10 @@ def _compile_vop3p(inst: ir3.VOP3P | ir4.VOP3P, ctx: _Ctx) -> UOp:
   lane = ctx.range()
   exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
   vdst_reg = ctx.inst_field(type(inst).vdst)
-  src0 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src0), lane, 16)
-  src1 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src1), lane, 16)
-  src2 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src2), lane, 16)
+  do_cast = any(x in op_name for x in ('F16', 'F32', 'BF16')) and 'IU' not in op_name
+  src0 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src0), lane, 16, do_cast=do_cast)
+  src1 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src1), lane, 16, do_cast=do_cast)
+  src2 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src2), lane, 16, do_cast=do_cast)
   opsel, opsel_hi = getattr(inst, 'opsel', 0) or 0, getattr(inst, 'opsel_hi', 3) if getattr(inst, 'opsel_hi', 3) is not None else 3
   opsel_hi2 = getattr(inst, 'opsel_hi2', 1) if getattr(inst, 'opsel_hi2', 1) is not None else 1
   neg, neg_hi = getattr(inst, 'neg', 0) or 0, getattr(inst, 'neg_hi', 0) or 0
@@ -1077,7 +1098,7 @@ def _get_runner(inst_bytes: bytes, arch: str = "rdna3"):
   canonical_name = f"{_op_name(inst).lower()}_{base.to_bytes(size, 'little').hex()}"
   sink = sink.replace(arg=KernelInfo(name=canonical_name)).rtag(1)
 
-  with Context(NOOPT=1, CHECK_OOB=0, TUPLE_ORDER=0, EMULATED_DTYPES=""):
+  with Context(NOOPT=1, CHECK_OOB=0, EMULATED_DTYPES=""):
     runner = get_runner('CPU', sink)
   _canonical_runner_cache.append((base, mask, size, runner))
   return runner, True
@@ -1120,8 +1141,8 @@ class WaveState:
     self.n_lanes = n_lanes
     self.vgpr_buf = Buffer('CPU', VGPR_SIZE, dtypes.uint32).ensure_allocated()
     self.sgpr_buf = Buffer('CPU', SGPR_COUNT, dtypes.uint32).ensure_allocated()
-    self._vgpr_mv = self.vgpr_buf.as_buffer(force_zero_copy=True).cast('I')
-    self._sgpr_mv = self.sgpr_buf.as_buffer(force_zero_copy=True).cast('I')
+    self._vgpr_mv = self.vgpr_buf.as_memoryview(force_zero_copy=True).cast('I')
+    self._sgpr_mv = self.sgpr_buf.as_memoryview(force_zero_copy=True).cast('I')
     # Zero memory using ctypes memset (much faster than Python loops)
     ctypes.memset(self.vgpr_buf._buf.va_addr, 0, VGPR_SIZE * 4)
     ctypes.memset(self.sgpr_buf._buf.va_addr, 0, SGPR_COUNT * 4)

extra/assembly/amd/pcode.py

@@ -440,7 +440,7 @@ class Parser:
       self.eat('COMMA')
       lo = self.parse()
       self.eat('RBRACE')
-      return (hi.cast(dtypes.uint64) << _u64(32)) | lo.cast(dtypes.uint64)
+      return (hi.cast(dt:=_BITS_DT.get((s:=lo.dtype.bitsize) * 2, dtypes.uint64)) << _const(dt, s)) | lo.cast(dt)
     if self.at('NUM'):
       num = self.eat('NUM').val
       if self.try_eat('QUOTE'):

extra/assembly/amd/test/hw/test_sop.py

@@ -87,6 +87,50 @@ class TestBasicScalar(unittest.TestCase):
     self.assertEqual(st.sgpr[2], f2i(2.0))
 
 
+class TestPack(unittest.TestCase):
+  """Tests for S_PACK instructions."""
+
+  def test_s_pack_ll_b32_b16(self):
+    """S_PACK_LL_B32_B16 packs low 16 bits of two sources into one 32-bit result."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADAAAA),
+      s_mov_b32(s[1], 0xDEADBBBB),
+      s_pack_ll_b32_b16(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xBBBBAAAA)
+
+  def test_s_pack_lh_b32_b16(self):
+    """S_PACK_LH_B32_B16: D0 = { S1[31:16], S0[15:0] }."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADAAAA),
+      s_mov_b32(s[1], 0xDEADBBBB),
+      s_pack_lh_b32_b16(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xDEADAAAA)
+
+  def test_s_pack_hh_b32_b16(self):
+    """S_PACK_HH_B32_B16: D0 = { S1[31:16], S0[31:16] }."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADAAAA),
+      s_mov_b32(s[1], 0xDEADBBBB),
+      s_pack_hh_b32_b16(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xDEADDEAD)
+
+  def test_s_pack_hl_b32_b16(self):
+    """S_PACK_HL_B32_B16: D0 = { S1[15:0], S0[31:16] }."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADAAAA),
+      s_mov_b32(s[1], 0xDEADBBBB),
+      s_pack_hl_b32_b16(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xBBBBDEAD)
+
+
 class TestQuadmaskWqm(unittest.TestCase):
   """Tests for S_QUADMASK_B32 and S_WQM_B32."""
 

extra/assembly/amd/test/hw/test_vop3.py

@@ -2949,6 +2949,306 @@ class TestVOP3Clamp(unittest.TestCase):
     self.assertAlmostEqual(i2f(st.vgpr[3][1]), 1.0, places=5, msg="lane 3: 2.5 should clamp to 1.0")
 
 
+class TestVOP3ClampUint32(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on unsigned 32-bit integer operations."""
+
+  def test_v_sub_nc_u32_e64_clamp_underflow(self):
+    """V_SUB_NC_U32_E64 with clamp: 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 1),
+      v_sub_nc_u32_e64(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_nc_u32_e64_clamp_no_underflow(self):
+    """V_SUB_NC_U32_E64 with clamp: 100 - 50 = 50 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 50),
+      v_sub_nc_u32_e64(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 50, f"expected 50, got {st.vgpr[0][2]}")
+
+  def test_v_add_nc_u32_e64_clamp_overflow(self):
+    """V_ADD_NC_U32_E64 with clamp: 0xFFFFFFFF + 1 should saturate to 0xFFFFFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[1], 1),
+      v_add_nc_u32_e64(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFF, f"expected 0xFFFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_nc_u32_e64_clamp_no_overflow(self):
+    """V_ADD_NC_U32_E64 with clamp: 100 + 50 = 150 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 50),
+      v_add_nc_u32_e64(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 150, f"expected 150, got {st.vgpr[0][2]}")
+
+
+class TestVOP3ClampUint16(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on unsigned 16-bit integer operations."""
+
+  def test_v_sub_nc_u16_clamp_underflow(self):
+    """V_SUB_NC_U16 with clamp: 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 1),
+      v_sub_nc_u16(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2] & 0xFFFF, 0, f"expected 0, got 0x{st.vgpr[0][2] & 0xFFFF:04x}")
+
+  def test_v_sub_nc_u16_clamp_no_underflow(self):
+    """V_SUB_NC_U16 with clamp: 100 - 50 = 50 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 50),
+      v_sub_nc_u16(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2] & 0xFFFF, 50, f"expected 50, got {st.vgpr[0][2] & 0xFFFF}")
+
+  def test_v_add_nc_u16_clamp_overflow(self):
+    """V_ADD_NC_U16 with clamp: 0xFFFF + 1 should saturate to 0xFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFF),
+      v_mov_b32_e32(v[1], 1),
+      v_add_nc_u16(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2] & 0xFFFF, 0xFFFF, f"expected 0xFFFF, got 0x{st.vgpr[0][2] & 0xFFFF:04x}")
+
+  def test_v_add_nc_u16_clamp_no_overflow(self):
+    """V_ADD_NC_U16 with clamp: 100 + 50 = 150 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 50),
+      v_add_nc_u16(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2] & 0xFFFF, 150, f"expected 150, got {st.vgpr[0][2] & 0xFFFF}")
+
+
+class TestVOP3ClampInt32(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on signed 32-bit integer operations."""
+
+  def test_v_add_nc_i32_clamp_overflow(self):
+    """V_ADD_NC_I32 with clamp: INT_MAX + 1 should saturate to INT_MAX."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x7FFFFFFF),  # S0 = INT_MAX
+      v_mov_b32_e32(v[1], 1),            # S1 = 1
+      v_add_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x7FFFFFFF, f"expected 0x7FFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_nc_i32_clamp_underflow(self):
+    """V_ADD_NC_I32 with clamp: INT_MIN + (-1) should saturate to INT_MIN."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x80000000),  # S0 = INT_MIN
+      v_mov_b32_e32(v[1], 0xFFFFFFFF),  # S1 = -1
+      v_add_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x80000000, f"expected 0x80000000, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_nc_i32_clamp_underflow(self):
+    """V_SUB_NC_I32 with clamp: INT_MIN - 1 should saturate to INT_MIN."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x80000000),  # S0 = INT_MIN
+      v_mov_b32_e32(v[1], 1),            # S1 = 1
+      v_sub_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x80000000, f"expected 0x80000000, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_nc_i32_clamp_overflow(self):
+    """V_SUB_NC_I32 with clamp: INT_MAX - (-1) should saturate to INT_MAX."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x7FFFFFFF),  # S0 = INT_MAX
+      v_mov_b32_e32(v[1], 0xFFFFFFFF),  # S1 = -1
+      v_sub_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x7FFFFFFF, f"expected 0x7FFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_nc_i32_no_saturation_positive(self):
+    """V_ADD_NC_I32 with clamp: 100 + 200 = 300 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 200),
+      v_add_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 300, f"expected 300, got {st.vgpr[0][2]}")
+
+  def test_v_add_nc_i32_no_saturation_negative(self):
+    """V_ADD_NC_I32 with clamp: -100 + -200 = -300 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFF9C),  # -100
+      v_mov_b32_e32(v[1], 0xFFFFFF38),  # -200
+      v_add_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    expected = 0xFFFFFED4  # -300
+    self.assertEqual(st.vgpr[0][2], expected, f"expected 0x{expected:08x}, got 0x{st.vgpr[0][2]:08x}")
+
+
+class TestVOP3ClampCarry(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on carry operations (VOP3SD)."""
+
+  def test_v_add_co_u32_clamp_overflow(self):
+    """V_ADD_CO_U32 with clamp: 0xFFFFFFFF + 1 should saturate to 0xFFFFFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[1], 1),
+      v_add_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFF, f"expected 0xFFFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_co_u32_clamp_no_overflow(self):
+    """V_ADD_CO_U32 with clamp: 100 + 200 = 300 (no saturation)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 200),
+      v_add_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 300, f"expected 300, got {st.vgpr[0][2]}")
+
+  def test_v_sub_co_u32_clamp_underflow(self):
+    """V_SUB_CO_U32 with clamp: 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 1),
+      v_sub_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_co_u32_clamp_no_underflow(self):
+    """V_SUB_CO_U32 with clamp: 300 - 100 = 200 (no saturation)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 300),
+      v_mov_b32_e32(v[1], 100),
+      v_sub_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 200, f"expected 200, got {st.vgpr[0][2]}")
+
+  def test_v_subrev_co_u32_clamp_underflow(self):
+    """V_SUBREV_CO_U32 with clamp: 1 - 0 reversed = 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1),  # This becomes the subtrahend
+      v_mov_b32_e32(v[1], 0),  # This becomes the minuend (0 - 1)
+      v_subrev_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_co_ci_u32_clamp_overflow(self):
+    """V_ADD_CO_CI_U32 with clamp: 0xFFFFFFFF + 1 + 0 should saturate to 0xFFFFFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[1], 1),
+      s_mov_b64(VCC, 0),  # No carry in
+      v_add_co_ci_u32(v[2], VCC, v[0], v[1], VCC, clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFF, f"expected 0xFFFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_co_ci_u32_clamp_overflow_with_carry(self):
+    """V_ADD_CO_CI_U32 with clamp: 0xFFFFFFFE + 1 + 1 should saturate to 0xFFFFFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFFFE),
+      v_mov_b32_e32(v[1], 1),
+      s_mov_b64(VCC, 1),  # Carry in = 1
+      v_add_co_ci_u32(v[2], VCC, v[0], v[1], VCC, clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFF, f"expected 0xFFFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_co_ci_u32_clamp_underflow(self):
+    """V_SUB_CO_CI_U32 with clamp: 0 - 1 - 0 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 1),
+      s_mov_b64(VCC, 0),  # No borrow in
+      v_sub_co_ci_u32(v[2], VCC, v[0], v[1], VCC, clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_subrev_co_ci_u32_clamp_underflow(self):
+    """V_SUBREV_CO_CI_U32 with clamp: reversed 1 - 0 - 0 = 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b64(VCC, 0),
+      v_subrev_co_ci_u32(v[2], VCC, v[0], v[1], VCC, clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+
+class TestVOP3ClampMAD(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on MAD (multiply-add) operations."""
+
+  def test_v_mad_u16_clamp_overflow(self):
+    """V_MAD_U16 with clamp: 0xFFFF * 2 + 0 should saturate to 0xFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFF),
+      v_mov_b32_e32(v[1], 2),
+      v_mov_b32_e32(v[2], 0),
+      v_mad_u16(v[3], v[0], v[1], v[2], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 0xFFFF, f"expected 0xFFFF, got 0x{st.vgpr[0][3] & 0xFFFF:04x}")
+
+  def test_v_mad_u16_clamp_overflow_with_add(self):
+    """V_MAD_U16 with clamp: 0x8000 * 2 + 0x1000 should saturate to 0xFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x8000),  # 32768
+      v_mov_b32_e32(v[1], 2),        # * 2 = 65536
+      v_mov_b32_e32(v[2], 0x1000),   # + 4096 = 69632 > 0xFFFF
+      v_mad_u16(v[3], v[0], v[1], v[2], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 0xFFFF, f"expected 0xFFFF, got 0x{st.vgpr[0][3] & 0xFFFF:04x}")
+
+  def test_v_mad_u16_no_overflow(self):
+    """V_MAD_U16 with clamp: 100 * 100 + 50 = 10050 (no saturation)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 100),
+      v_mov_b32_e32(v[2], 50),
+      v_mad_u16(v[3], v[0], v[1], v[2], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 10050, f"expected 10050, got {st.vgpr[0][3] & 0xFFFF}")
+
+  def test_v_mad_u16_no_clamp(self):
+    """V_MAD_U16 without clamp: 0xFFFF * 2 + 0 should wrap to 0xFFFE."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFF),
+      v_mov_b32_e32(v[1], 2),
+      v_mov_b32_e32(v[2], 0),
+      v_mad_u16(v[3], v[0], v[1], v[2], clmp=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # 0xFFFF * 2 = 0x1FFFE, low 16 bits = 0xFFFE
+    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 0xFFFE, f"expected 0xFFFE, got 0x{st.vgpr[0][3] & 0xFFFF:04x}")
+
+
 class TestCvtPkF16(unittest.TestCase):
   """Tests for V_CVT_PK_RTZ_F16_F32 - pack two f32 to f16 with round toward zero."""
 

extra/assembly/amd/test/hw/test_vop3p.py

@@ -390,6 +390,24 @@ class TestVOP3P(unittest.TestCase):
     self.assertAlmostEqual(lo, 6.0, places=1)
     self.assertAlmostEqual(hi, 0.0, places=1)
 
+  def test_v_pk_add_u16_float_inline_const_opsel(self):
+    """V_PK_ADD_U16 with float inline constant 2.0
+    Regression test: for integer packed ops, do not perform the f32->f16 conversion.
+    """
+    # src1 = inline float constant 2.0
+    instructions = [
+      s_mov_b32(s[0], 0x00030005),  # packed u16: hi=3, lo=5
+      v_mov_b32_e32(v[0], s[0]),
+      v_pk_add_u16(v[1], v[0], SrcEnum.POS_TWO, opsel_hi=3, opsel_hi2=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1]
+    lo = result & 0xffff
+    hi = (result >> 16) & 0xffff
+    # lo = 5 + 0x0000 = 0x0005, hi = 3 + 0x4000 = 0x4003
+    self.assertEqual(lo, 0x0005, f"lo: expected 0x0005, got 0x{lo:04x}")
+    self.assertEqual(hi, 0x4003, f"hi: expected 0x4003, got 0x{hi:04x}")
+
 
 class TestWMMAF16(unittest.TestCase):
   """Tests for WMMA F16 output variant (V_WMMA_F16_16X16X16_F16).

extra/fp8/fp8_linear.py

@@ -28,7 +28,7 @@ def custom_matmul(output: UOp, inp: UOp, weight: UOp) -> UOp:
   return store_op.sink(arg=KernelInfo(name=f"fp8_matmul_{inp.shape}x{weight.shape}"))
 
 def custom_matmul_backward(gradient: UOp, kernel: UOp) -> tuple[UOp, UOp]:
-  _, input_uop, weight_uop = kernel.src
+  _, input_uop, weight_uop = kernel.src[1:]
   input_tensor = Tensor(input_uop, device=input_uop.device)
   grad_tensor = Tensor(gradient, device=gradient.device)
   weight_tensor = Tensor(weight_uop, device=weight_uop.device)

extra/gemm/asm/cdna/asm.py

@@ -9,11 +9,17 @@ GEMM_ARGS = {
   (8192, 4096, 4096): (256, 64, 32768),
   (8192, 14336, 4096): (256, 64, 114688),
   (8192, 4096, 14336): (256, 224, 114688),
-  (8192, 128256, 4096): (16032, 64, 1026048),
+  # TODO: get a fast gemm for this shape
+  #(8192, 128256, 4096): (16032, 64, 1026048),
   (8192, 8192, 8192): (256, 128, 131072),
   (4096, 4096, 4096): (256, 64, 16384),
   (4096, 14336, 4096): (256, 64, 57344),
+  (4096, 14336, 8192): (256, 128, 114688),
   (4096, 4096, 14336): (256, 224, 57344),
+  (14336, 4096, 8192): (256, 128, 114688),
+  (4096, 8192, 14336): (256, 224, 114688),
+  (4096, 4096, 8192): (256, 128, 32768),
+  (4096, 8192, 4096): (256, 64, 32768),
 }
 ITERS_ARGS = {64: (67108864, 0), 128: (33554432, 0), 224: (613566757, 2147483656)}
 

extra/gemm/asm/cdna/gemm.py

@@ -10,6 +10,7 @@ from extra.gemm.asm.cdna.asm import build_kernel, GEMM_ARGS
 
 WORKGROUP_SIZE = 256
 
+@functools.cache
 def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str, arch:str, wg:int) -> UOp:
   batch, M, K = A.shape
   K2, N = B.shape[(1 if B.ndim == 3 else 0):]
@@ -19,6 +20,7 @@ def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str, arch:str, wg:int) -> UOp:
   k = build_kernel(batch, M, N, K, A.dtype.base)
   sink = UOp.sink(C.base, A.base, B.base, lidx, gidx,
                   arg=KernelInfo(name=k.name, estimates=Estimates(ops=2*batch*M*N*K, mem=(batch*M*K + K*N + batch*M*N)*2)))
+  # TODO: you shouldn't have to call the compiler here, BINARY should be auto-added
   binary = HIPCompiler(arch).compile(k.to_asm())
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                UOp(Ops.SOURCE, arg=k.to_text()), UOp(Ops.BINARY, arg=binary)))
@@ -30,14 +32,18 @@ atexit.register(lambda: print(f'asm_gemm: {counters["used"]} used, {len(counters
 def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
   if a.dtype != b.dtype: return todo(f"dtypes must match {a.dtype} != {b.dtype}")
   if a.dtype not in {dtypes.bfloat16, dtypes.float16}: return todo(f"only bfloat16/float16, got {a.dtype}")
-  # only sharding on the batch is tested, others might work too
-  if isinstance(a.device, tuple) and not (a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None):
-    return todo(f"sharding mismatch a.ndim={a.ndim} a.uop.axis={a.uop.axis} b.uop.axis={b.uop.axis}")
   batch, M, K = (1, *a.shape) if a.ndim == 2 else a.shape
   N = b.shape[1]
-  if isinstance(a.device, tuple): batch //= len(a.device)
+  # only sharding on the batch or K is tested, others might work too
+  if isinstance(a.device, tuple):
+    if a.ndim == 2 and a.uop.axis == 1 and b.uop.axis == 0: K //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None: batch //= len(a.device)
+    else: return todo(f"sharding mismatch a.ndim={a.ndim} a.uop.axis={a.uop.axis} b.uop.axis={b.uop.axis}")
+    dname = a.device[0]
+  else: dname = a.device
+  arch = getattr(Device[dname].renderer, "arch", "")
   if batch not in {1, 2}: return todo(f"GEMM batch size {batch}")
-  if (key:=(M, N, K)) not in GEMM_ARGS: return todo(f"GEMM shape not supported {key}")
+  if (key:=(M, N, K)) not in GEMM_ARGS and arch == "gfx950": return todo(f"GEMM shape not supported {key} on {arch}")
   return True
 
 # ** UOp gemm to test Tensor.custom_kernel multi and backward correctness on non cdna4
@@ -58,14 +64,13 @@ def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
 # ** backward gemm, might use the asm gemm
 
 def custom_gemm_bw(gradient:UOp, kernel:UOp):
-  out, a, b = kernel.src
+  out, a, b = kernel.src[1:]
   assert all_same([gradient.device, a.device, b.device, out.device])
   a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
+  # TODO: this needs to be cleaned up and done properly, the batch dim of grad and a multi need to align
+  g_t = g_t[:a.shape[0]]
   grad_a = (g_t @ b_t.T).uop
-  a_T = a_t.transpose(-2, -1)
-  a_T = a_T.reshape(*a_T.shape[:-1], 1, a_T.shape[-1])
-  g_r = g_t.reshape(*g_t.shape[:-2], 1, *g_t.shape[-2:]).transpose(-1, -2)
-  grad_b = (a_T * g_r).sum((-1, 0)).uop
+  grad_b = (a_t.permute(2, 0, 1).reshape(a_t.shape[2], -1) @ g_t.reshape(-1, g_t.shape[-1])).uop
   return (None, grad_a, grad_b)
 
 # ** main gemm function
@@ -79,17 +84,19 @@ def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
   batch, M, K = a.shape
   N = b.shape[1]
   is_multi = isinstance(a.device, tuple)
+  if (k_sharded:=is_multi and a.uop.axis == 2): K //= len(a.device)
 
   if is_multi:
-    out = Tensor(Tensor.empty(batch//len(a.device), M, N, dtype=a.dtype, device=a.device).uop.multi(0), device=a.device)
+    out = Tensor(Tensor.empty(batch//len(a.device) if a.uop.axis==0 else batch, M, N, dtype=a.dtype, device=a.device).uop.multi(0), device=a.device)
   else:
     out = Tensor.empty(batch, M, N, dtype=a.dtype, device=a.device)
 
   dname = a.device[0] if is_multi else a.device
-  arch = getattr(Device[dname].renderer, "arch", None)
+  arch = getattr(Device[dname].renderer, "arch", "")
   if arch.startswith("gfx950") and getenv("USE_ASM", 1):
     numWG = GEMM_ARGS[(M, N, K)][0]
     out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname, wg=numWG, arch=arch), grad_fxn=custom_gemm_bw)[0]
   else:
     out = Tensor.custom_kernel(out, a, b, fxn=custom_uop_gemm, grad_fxn=custom_gemm_bw)[0]
+  if k_sharded: out = out.sum(0)
   return out.squeeze(0) if squeeze else out

extra/gemm/asm/cdna/test_asm_gemm.py

@@ -1,46 +0,0 @@
-import unittest
-from tinygrad import Tensor, Device, dtypes, Context
-from tinygrad.helpers import getenv
-from extra.gemm.asm.cdna.gemm import asm_gemm
-
-def verify_asm_gemm(batch:int, M:int, N:int, K:int, dtype=dtypes.bfloat16, multi=False) -> None:
-  Tensor.manual_seed(0)
-  a_rand = Tensor.randn((batch, M, K), dtype=dtypes.float).sub(0.5).cast(dtype)
-  b_rand = Tensor.randn((K, N), dtype=dtypes.float).sub(0.5).cast(dtype)
-  with Context(DEBUG=0):
-    Tensor.realize(a_rand, b_rand)
-
-  devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(8)) if multi else None
-
-  a, b = Tensor(a_rand.numpy(), requires_grad=True).cast(dtype), Tensor(b_rand.numpy(), requires_grad=True).cast(dtype)
-  if multi: a, b = a.shard(devs, axis=0), b.shard(devs, axis=None)
-  tst = asm_gemm(a, b)
-  tst.sum().backward()
-  Tensor.realize(tst, a.grad, b.grad)
-
-  a_ref, b_ref = Tensor(a_rand.numpy(), requires_grad=True).cast(dtype), Tensor(b_rand.numpy(), requires_grad=True).cast(dtype)
-  if multi: a_ref, b_ref = a_ref.shard(devs, axis=0), b_ref.shard(devs, axis=None)
-  with Context(ASM_GEMM=0): ref = a_ref @ b_ref
-  ref.sum().backward()
-  Tensor.realize(ref, a_ref.grad, b_ref.grad)
-
-  with Context(DEBUG=0):
-    assert (tst - ref).square().max().float().item() < 1e-6, "forward mismatch"
-    assert (a.grad - a_ref.grad).square().max().float().item() < 1e-3, "grad_a mismatch"
-    assert (b.grad - b_ref.grad).square().max().float().item() < 1e-3, "grad_b mismatch"
-
-class TestGemm(unittest.TestCase):
-  def test_simple(self): verify_asm_gemm(1, N:=getenv("N", 4096), N, N, dtype=dtypes.half)
-
-  def test_gemm1(self): verify_asm_gemm(8, 8192, 4096, 14336, multi=True)
-  def test_gemm2(self): verify_asm_gemm(8, 8192, 128256, 4096, multi=True)
-  def test_gemm3(self): verify_asm_gemm(8, 8192, 14336, 4096, multi=True)
-  def test_gemm4(self): verify_asm_gemm(8, 4096, 14336, 4096, multi=True)
-  def test_gemm5(self): verify_asm_gemm(8, 4096, 4096, 14336, multi=True)
-  def test_gemm6(self): verify_asm_gemm(16, 4096, 4096, 14336, multi=True)
-  def test_gemm_unsupported(self):
-    with self.assertRaisesRegex(AssertionError, "shape not supported"):
-      verify_asm_gemm(8, 8192, 1024, 4096, multi=True)
-
-if __name__ == "__main__":
-  unittest.main()

extra/gemm/intel_xmx.py

@@ -37,7 +37,7 @@ b.copyin(row.data)
 c.copyin(mat.data)
 ret = prog(a._buf, b._buf, c._buf, global_size=[1,1,1], local_size=[8,1,1], wait=True)
 print(ret)
-out = np.frombuffer(a.as_buffer(), np.float32)
+out = np.frombuffer(a.as_memoryview(), np.float32)
 real = row.astype(np.float32)@mat.T.astype(np.float32)
 print("out:", out)
 print("real", real)

extra/gemm/triton_nv_matmul.py

@@ -98,10 +98,10 @@ if __name__ == "__main__":
   # check correctness
   if getenv("VERIFY"):
     from tinygrad.engine.realize import run_schedule
-    triton_buf = np.frombuffer(si.bufs[0].as_buffer(), np.float16).reshape(M,N)
+    triton_buf = np.frombuffer(si.bufs[0].as_memoryview(), np.float16).reshape(M,N)
     print(triton_buf)
     run_schedule(sched)
-    tinygrad_buf = np.frombuffer(si.bufs[0].as_buffer(), np.float16).reshape(M,N)
+    tinygrad_buf = np.frombuffer(si.bufs[0].as_memoryview(), np.float16).reshape(M,N)
     print(tinygrad_buf)
     np.testing.assert_allclose(triton_buf, tinygrad_buf)
     print("correct!")

extra/hip_large_kernel.py

@@ -18,7 +18,7 @@ prg = dev.runtime("write_ones", mbin)
 prg(buf0._buf, global_size=(1,65537,1), local_size=(1,1,1), wait=True)
 
 import numpy as np
-def to_np(buf): return np.frombuffer(buf.as_buffer().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))
+def to_np(buf): return np.frombuffer(buf.as_memoryview().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))
 
 big = to_np(buf0)
 print(big)

extra/huggingface_onnx/huggingface_manager.py

@@ -8,14 +8,14 @@ from tinygrad.helpers import _ensure_downloads_dir
 DOWNLOADS_DIR = _ensure_downloads_dir() / "models"
 from tinygrad.helpers import tqdm
 
-def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, cache_dir: str|Path|None=None,
+def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, local_dir: str|Path|None=None,
                                  tries: int=2, **kwargs) -> Path:
   for attempt in range(tries):
     try:
       return Path(snapshot_download(
         repo_id=repo_id,
         allow_patterns=allow_patterns,
-        cache_dir=str(cache_dir) if cache_dir is not None else None,
+        local_dir=str(local_dir) if local_dir is not None else None,
         **kwargs
       ))
     except Exception as e:
@@ -144,14 +144,14 @@ class HuggingFaceONNXManager:
         root_path = snapshot_download_with_retry(
           repo_id=model_id,
           allow_patterns=allow_patterns,
-          cache_dir=str(self.models_dir)
+          local_dir=str(self.models_dir / model_id)
         )
 
         # Download config files (usually small)
         snapshot_download_with_retry(
           repo_id=model_id,
           allow_patterns=["*config.json"],
-          cache_dir=str(self.models_dir)
+          local_dir=str(self.models_dir / model_id)
         )
 
         model_data["download_path"] = str(root_path)

extra/huggingface_onnx/run_models.py

@@ -88,8 +88,8 @@ if __name__ == "__main__":
       # repo id
       # validates all onnx models inside repo
       repo_id = "/".join(path)
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], cache_dir=DOWNLOADS_DIR)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], local_dir=DOWNLOADS_DIR / repo_id)
+      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], local_dir=DOWNLOADS_DIR / repo_id)
       config = get_config(root_path)
       for onnx_model in root_path.rglob("*.onnx"):
         rtol, atol = get_tolerances(onnx_model.name)
@@ -101,8 +101,8 @@ if __name__ == "__main__":
       onnx_model = path[-1]
       assert path[-1].endswith(".onnx")
       repo_id, relative_path = "/".join(path[:2]), "/".join(path[2:])
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], cache_dir=DOWNLOADS_DIR)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], local_dir=DOWNLOADS_DIR / repo_id)
+      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], local_dir=DOWNLOADS_DIR / repo_id)
       config = get_config(root_path)
       rtol, atol = get_tolerances(onnx_model)
       print(f"validating {relative_path} with truncate={args.truncate}, {rtol=}, {atol=}")

extra/models/clip.py

@@ -5,7 +5,6 @@ from tinygrad.nn import Linear, LayerNorm, Embedding, Conv2d
 from typing import List, Optional, Union, Tuple, Dict
 from abc import ABC, abstractmethod
 from functools import lru_cache
-from PIL import Image
 import numpy as np
 import re, gzip
 
@@ -444,7 +443,8 @@ class OpenClipEncoder:
   # TODO:
   # Should be doable in pure tinygrad, would just require some work and verification.
   # This is very desirable since it would allow for full generation->evaluation in a single JIT call.
-  def prepare_image(self, image:Image.Image) -> Tensor:
+  def prepare_image(self, image) -> Tensor:
+    from PIL import Image
     SIZE = 224
     w, h = image.size
     scale = min(SIZE / h, SIZE / w)

extra/models/llama.py

@@ -55,12 +55,15 @@ class Attention:
       xqkv = x @ self.wqkv.T
       xq, xk, xv = xqkv.split([self.wq.weight.shape[0], self.wk.weight.shape[0], self.wv.weight.shape[0]], dim=2)
     else:
-      xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
+      xq, xk, xv = self.wq(x), self.wk(x.contiguous_backward()), self.wv(x)
 
     if self.q_norm is not None and self.k_norm is not None:
       xq = self.q_norm(xq)
       xk = self.k_norm(xk)
 
+    # cast_float_to_bf16 is expensive in reduction loops, break it out
+    if x.dtype == dtypes.bfloat16: xq, xk = xq.contiguous_backward(), xk.contiguous_backward()
+
     xq = xq.reshape(xq.shape[0], xq.shape[1], self.n_heads, self.head_dim)
     xk = xk.reshape(xk.shape[0], xk.shape[1], self.n_kv_heads, self.head_dim)
     xv = xv.reshape(xv.shape[0], xv.shape[1], self.n_kv_heads, self.head_dim)
@@ -86,20 +89,23 @@ class Attention:
       assert start_pos == 0
       keys, values = xk, xv
 
-    keys, values = repeat_kv(keys, self.n_rep), repeat_kv(values, self.n_rep)
-    xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2)
+    if self.max_context:
+      keys, values = repeat_kv(keys, self.n_rep), repeat_kv(values, self.n_rep)
+      xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
+      attn = xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2)
+    else:
+      xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
+      attn = xq.scaled_dot_product_attention(keys, values, is_causal=True, enable_gqa=True).transpose(1, 2)
     if getenv("STUB_ATTENTION"):
-      # TODO: do we need mask?
       from tinygrad.uop.ops import UOp, KernelInfo
       def fa_custom_forward(attn:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
         return UOp.sink(arg=KernelInfo(name="fa_custom_forward"))
       def fa_custom_backward(out_q:UOp, out_k:UOp, out_v:UOp, grad:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
         return UOp.sink(arg=KernelInfo(name="fa_custom_backward"))
       def fa_backward(grad:UOp, kernel:UOp) -> tuple[None, UOp, UOp, UOp]:
-        grad_q = Tensor.empty_like(q:=Tensor(kernel.src[1]))
-        grad_k = Tensor.empty_like(k:=Tensor(kernel.src[2]))
-        grad_v = Tensor.empty_like(v:=Tensor(kernel.src[3]))
+        grad_q = Tensor.empty_like(q:=Tensor(kernel.src[2]))
+        grad_k = Tensor.empty_like(k:=Tensor(kernel.src[3]))
+        grad_v = Tensor.empty_like(v:=Tensor(kernel.src[4]))
         ck = Tensor.custom_kernel(grad_q, grad_k, grad_v, Tensor(grad), q, k, v, fxn=fa_custom_backward)[:3]
         return (None, ck[0].uop, ck[1].uop, ck[2].uop)
       attn = Tensor.empty_like(attn).custom_kernel(xq, keys, values, fxn=fa_custom_forward, grad_fxn=fa_backward)[0]
@@ -197,7 +203,9 @@ class Transformer:
     h = self.tok_embeddings(tokens)
     freqs_cis = self.freqs_cis.cast(h.dtype)[:, start_pos:start_pos+seqlen, :, :, :]
 
-    mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype, device=h.device).triu(start_pos+1) if seqlen > 1 else None
+    if self.max_context != 0 and seqlen > 1:
+      mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype, device=h.device).triu(start_pos+1)
+    else: mask = None
     for layer in self.layers: h = layer(h, start_pos, freqs_cis, mask)
     logits = self.output(self.norm(h))
     if math.isnan(temperature): return logits

extra/optimization/generate_dataset.sh

@@ -7,7 +7,7 @@ export CAPTURE_PROCESS_REPLAY=1
 rm "$LOGOPS" 2>/dev/null || true
 test/external/process_replay/reset.py
 
-CI=1 python3 -m pytest -n=auto test/test_ops.py test/test_nn.py test/unit/test_winograd.py test/models/test_real_world.py --durations=20
+CI=1 python3 -m pytest -n=auto test/test_ops.py test/test_nn.py test/unit/test_winograd.py test/null/test_real_world.py --durations=20
 CL=1 python3 -m pytest test/test_tiny.py
 
 # extract, sort and uniq

extra/perfetto/to_perfetto.py

@@ -2,27 +2,26 @@ import sys, pickle, decimal, json
 from tinygrad.device import ProfileDeviceEvent, ProfileGraphEvent
 from tinygrad.helpers import tqdm, temp, ProfileEvent, ProfileRangeEvent, TracingKey
 
-devices:dict[str, tuple[decimal.Decimal, decimal.Decimal, int]] = {}
-def prep_ts(device:str, ts:decimal.Decimal, is_copy): return int(decimal.Decimal(ts) + devices[device][is_copy])
-def dev_to_pid(device:str, is_copy=False): return {"pid": devices[device][2], "tid": int(is_copy)}
+devices:dict[str, tuple[decimal.Decimal, int]] = {}
+def prep_ts(device:str, ts:decimal.Decimal): return int(decimal.Decimal(ts) + devices[device][0])
+def dev_to_pid(device:str): return {"pid": devices[device][1], "tid": 0}
 def dev_ev_to_perfetto_json(ev:ProfileDeviceEvent):
-  devices[ev.device] = (ev.comp_tdiff, ev.copy_tdiff if ev.copy_tdiff is not None else ev.comp_tdiff, len(devices))
+  devices[ev.device] = (ev.tdiff, len(devices))
   return [{"name": "process_name", "ph": "M", "pid": dev_to_pid(ev.device)['pid'], "args": {"name": ev.device}},
-          {"name": "thread_name", "ph": "M", "pid": dev_to_pid(ev.device)['pid'], "tid": 0, "args": {"name": "COMPUTE"}},
-          {"name": "thread_name", "ph": "M", "pid": dev_to_pid(ev.device)['pid'], "tid": 1, "args": {"name": "COPY"}}]
+          {"name": "thread_name", "ph": "M", "pid": dev_to_pid(ev.device)['pid'], "tid": 0, "args": {"name": ev.device}}]
 def range_ev_to_perfetto_json(ev:ProfileRangeEvent):
   name = ev.name.display_name if isinstance(ev.name, TracingKey) else ev.name
-  return [{"name": name, "ph": "X", "ts": prep_ts(ev.device, ev.st, ev.is_copy), "dur": float(ev.en-ev.st), **dev_to_pid(ev.device, ev.is_copy)}]
+  return [{"name": name, "ph": "X", "ts": prep_ts(ev.device, ev.st), "dur": float(ev.en-ev.st), **dev_to_pid(ev.device)}]
 def graph_ev_to_perfetto_json(ev:ProfileGraphEvent, reccnt):
   ret = []
   for i,e in enumerate(ev.ents):
     st, en = ev.sigs[e.st_id], ev.sigs[e.en_id]
     name = e.name.display_name if isinstance(e.name, TracingKey) else e.name
-    ret += [{"name": name, "ph": "X", "ts": prep_ts(e.device, st, e.is_copy), "dur": float(en-st), **dev_to_pid(e.device, e.is_copy)}]
+    ret += [{"name": name, "ph": "X", "ts": prep_ts(e.device, st), "dur": float(en-st), **dev_to_pid(e.device)}]
     for dep in ev.deps[i]:
       d = ev.ents[dep]
-      ret += [{"ph": "s", **dev_to_pid(d.device, d.is_copy), "id": reccnt+len(ret), "ts": prep_ts(d.device, ev.sigs[d.en_id], d.is_copy), "bp": "e"}]
-      ret += [{"ph": "f", **dev_to_pid(e.device, e.is_copy), "id": reccnt+len(ret)-1, "ts": prep_ts(e.device, st, e.is_copy), "bp": "e"}]
+      ret += [{"ph": "s", **dev_to_pid(d.device), "id": reccnt+len(ret), "ts": prep_ts(d.device, ev.sigs[d.en_id]), "bp": "e"}]
+      ret += [{"ph": "f", **dev_to_pid(e.device), "id": reccnt+len(ret)-1, "ts": prep_ts(e.device, st), "bp": "e"}]
   return ret
 def to_perfetto(profile:list[ProfileEvent]):
   # Start json with devices.

extra/sqtt/rgptool.py

@@ -145,7 +145,8 @@ class RGP:
   @staticmethod
   def from_profile(profile_pickled, device:str|None=None):
     profile: list[ProfileEvent] = pickle.loads(profile_pickled)
-    device_events = {x.device:x for x in profile if isinstance(x, ProfileDeviceEvent) and x.device.startswith('AMD')}
+    def _is_base_dev(d): return all(p.isdigit() for p in d.split(":")[1:])
+    device_events = {x.device:x for x in profile if isinstance(x, ProfileDeviceEvent) and x.device.startswith('AMD') and _is_base_dev(x.device)}
     if device is None:
       if len(device_events) == 0: raise RuntimeError('No supported devices found in profile')
       if len(device_events) > 1: raise RuntimeError(f"More than one supported device found, select which one to export: {', '.join(device_events.keys())}")

extra/thunder/tiny/fa.py

@@ -55,9 +55,6 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
       batch = ker.blockIdx_z
       q_seq = ker.blockIdx_y * NUM_WORKERS + ker.warpid
 
-      k_smem = ker.st((KV_BLOCK_SIZE, D), dtypes.bfloat16)
-      v_smem = ker.st((KV_BLOCK_SIZE, D), dtypes.bfloat16)
-
       q_reg_fl = ker.rt((Q_BLOCK_SIZE, D), dtypes.float32)
       q_reg = ker.rt((Q_BLOCK_SIZE, D), dtypes.bfloat16)
       q_reg_transposed = ker.rt((D, Q_BLOCK_SIZE), dtypes.bfloat16, TileLayout.COL)
@@ -89,11 +86,8 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
 
       num_kv_blocks = (q_seq + 1) if is_causal else (N // KV_BLOCK_SIZE)
       for kv_idx in ker.range(num_kv_blocks):
-        k_smem = warp.load(k_smem, k, (), (batch, kv_idx, head_kv, 0), axis=1)
-        v_smem = warp.load(v_smem, v, (), (batch, kv_idx, head_kv, 0), axis=1)
-
-        k_reg = warp.load(k_reg, k_smem)
-        v_reg = warp.load(v_reg, v_smem)
+        k_reg = warp.load(k_reg, k, (), (batch, kv_idx, head_kv, 0), axis=1)
+        v_reg = warp.load(v_reg, v, (), (batch, kv_idx, head_kv, 0), axis=1)
 
         # mma qk^t
         att_block = warp.zero(att_block.after(kv_idx))
@@ -169,9 +163,6 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
       batch = ker.blockIdx_z
       q_seq = ker.blockIdx_y * NUM_WORKERS + ker.warpid
 
-      k_smem = ker.st((KV_BLOCK_SIZE, D), dtypes.bfloat16)
-      v_smem = ker.st((KV_BLOCK_SIZE, D), dtypes.bfloat16)
-
       q_reg_fl = ker.rt((Q_BLOCK_SIZE, D), dtypes.float32)
       q_reg = ker.rt((Q_BLOCK_SIZE, D), dtypes.bfloat16)
       q_reg_t = ker.rt((D, Q_BLOCK_SIZE), dtypes.bfloat16, TileLayout.COL)
@@ -212,14 +203,12 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
 
       num_kv_blocks = (q_seq + 1) if is_causal else (N // KV_BLOCK_SIZE)
       for kv_idx in ker.range(num_kv_blocks):
-        k_smem = warp.load(k_smem, k, (), (batch, kv_idx, head_kv, 0), axis=1)
-        v_smem = warp.load(v_smem, v, (), (batch, kv_idx, head_kv, 0), axis=1)
+        k_reg = warp.load(k_reg, k, (), (batch, kv_idx, head_kv, 0), axis=1)
+        k_reg_col = warp.load(k_reg_col, k, (), (batch, kv_idx, head_kv, 0), axis=1)
+        v_reg = warp.load(v_reg, v, (), (batch, kv_idx, head_kv, 0), axis=1)
 
-        k_reg = warp.load(k_reg, k_smem)
         k_reg_t = warp.transpose(k_reg_t, k_reg)
-        k_reg_col = warp.load(k_reg_col, k_smem)
         k_reg_col_t = warp.transpose(k_reg_col_t, k_reg_col)
-        v_reg = warp.load(v_reg, v_smem)
 
         # mma qk^t
         att_block = warp.zero(att_block.after(kv_idx))
@@ -273,8 +262,6 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
       batch = ker.blockIdx_z
       kv_seq = ker.blockIdx_y * NUM_WORKERS + ker.warpid
 
-      q_smem = ker.st((Q_BLOCK_SIZE, D), dtypes.bfloat16)
-      do_smem = ker.st((Q_BLOCK_SIZE, D), dtypes.bfloat16)
       att_smem = ker.st((Q_BLOCK_SIZE, KV_BLOCK_SIZE), dtypes.bfloat16)
 
       q_reg = ker.rt((Q_BLOCK_SIZE, D), dtypes.bfloat16)
@@ -308,19 +295,17 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
       k_reg_t = warp.transpose(k_reg_t, k_reg)
       v_reg = warp.load(v_reg, v, (), (batch, kv_seq, head_kv, 0), axis=1)
 
-      for q_idx in ker.range(N // Q_BLOCK_SIZE):
+      q_start = kv_seq if is_causal else 0
+      for q_idx in ker.range(q_start, N // Q_BLOCK_SIZE):
         for g in ker.range(GROUP_SIZE):
           head_q = head_kv * GROUP_SIZE + g
 
-          # load q and do
-          q_smem = warp.load(q_smem, q, (), (batch, q_idx, head_q, 0), axis=1)
-          do_smem = warp.load(do_smem, do, (), (batch, q_idx, head_q, 0), axis=1)
+          q_reg = warp.load(q_reg, q, (), (batch, q_idx, head_q, 0), axis=1)
+          q_reg_col = warp.load(q_reg_col, q, (), (batch, q_idx, head_q, 0), axis=1)
+          do_reg = warp.load(do_reg, do, (), (batch, q_idx, head_q, 0), axis=1)
+          do_reg_col = warp.load(do_reg_col, do, (), (batch, q_idx, head_q, 0), axis=1)
 
-          q_reg = warp.load(q_reg, q_smem)
           q_reg_t = warp.transpose(q_reg_t, q_reg)
-          q_reg_col = warp.load(q_reg_col, q_smem)
-          do_reg = warp.load(do_reg, do_smem)
-          do_reg_col = warp.load(do_reg_col, do_smem)
 
           # load l_vec and delta_vec
           l_vec_reg = warp.load(l_vec_reg, l_vec, (), (batch, head_q, 0, q_idx), axis=2)

extra/thunder/tiny/tk/kernel.py

@@ -12,7 +12,7 @@ class _tk_range:
   def __next__(self):
     if not self.done:
       self.done = True
-      self._rng = UOp.range(self.end // self.step, self.rid, axis_type=self.axis_type) * self.step + self.start
+      self._rng = UOp.range((self.end - self.start) // self.step, self.rid, axis_type=self.axis_type) * self.step + self.start
       return self._rng
     raise StopIteration
 

extra/tinyfs/fetch_file.py

@@ -1,11 +1,36 @@
 from tinygrad.tensor import Tensor
-import argparse
+import argparse, math, hashlib
+
+def _python_hash_1mb(data:bytes|bytearray):
+  chunks = [data[i:i+4096] for i in range(0, len(data), 4096)]
+  chunk_hashes = [hashlib.shake_128(chunk).digest(16) for chunk in chunks]
+  return hashlib.shake_128(b''.join(chunk_hashes)).digest(16)
+
+def hash_file(data: bytes|bytearray):
+  if len(data) % Tensor.CHUNK_SIZE != 0: data += bytes(Tensor.CHUNK_SIZE - len(data) % Tensor.CHUNK_SIZE)
+  base_chunks = math.ceil(len(data) / Tensor.CHUNK_SIZE)
+  tree_depth = math.ceil(math.log(base_chunks, Tensor.CHUNK_SIZE // 16))
+
+  for _ in range(tree_depth + 1):
+    data_chunks = [data[i:i+Tensor.CHUNK_SIZE] for i in range(0, len(data), Tensor.CHUNK_SIZE)]
+    data_chunk_hashes = [_python_hash_1mb(chunk) for chunk in data_chunks]
+    data = b''.join(data_chunk_hashes)
+    if len(data) % Tensor.CHUNK_SIZE != 0: data += bytes(Tensor.CHUNK_SIZE - len(data) % Tensor.CHUNK_SIZE)
+
+  return data[:16]
 
 if __name__ == "__main__":
   parser = argparse.ArgumentParser()
   parser.add_argument("--hash", type=str, required=True, help="file hash to fetch")
   parser.add_argument("--len", type=int, required=True, help="file length to fetch")
   parser.add_argument("--dest", type=str, required=True, help="destination path to save the file")
+  parser.add_argument("--check", action="store_true", help="verify the file hash after fetching")
   args = parser.parse_args()
 
   Tensor(bytes.fromhex(args.hash), device="CPU").fs_load(args.len).to(f"disk:{args.dest}").realize()
+
+  if args.check:
+    with open(args.dest, "rb") as f:
+      data = f.read()
+      assert hash_file(data) == bytes.fromhex(args.hash), "Hash mismatch after fetching file"
+    print("File hash verified successfully!")

extra/torch_backend/backend.py

@@ -29,6 +29,10 @@ def wrap(x: Tensor) -> torch.Tensor:
   x._strides = strides_for_shape(x.shape) # always recalculate
   if (not hasattr(x, '_storage_offset')) or (not x.uop.is_realized): x._storage_offset = calculate_storage_offset(x)
   return mod.wrap(x, _to_torch_dtype(x.dtype), _to_torch_device(x.device).index)
+def _update_torch_metadata(tensor: torch.Tensor, tiny: Tensor) -> None:
+  tiny._strides = strides_for_shape(tiny.shape)
+  tiny._storage_offset = calculate_storage_offset(tiny)
+  mod.update_metadata(tensor, tiny.shape, tiny._strides, tiny._storage_offset)
 def unwrap(x:torch.Tensor) -> Tensor:
   assert isinstance(x, torch.Tensor), f"x isn't {type(x)}"
   return mod.unwrap(x)
@@ -344,7 +348,7 @@ def scatter_add(self, dim, index, src, out):
 def _copy_between_devices(src, dest, cast_dtype, to_device, non_blocking=False):
   if src.is_tiny and dest.is_tiny:
     src_t, dest_t = unwrap(src), unwrap(dest)
-    if dest_t.uop.is_contiguous() or dest_t.uop.is_realized: src_t = src_t.contiguous()
+    if dest_t.uop.has_buffer_identity() or dest_t.uop.is_realized: src_t = src_t.contiguous()
     _apply_inplace(dest_t, src_t.cast(cast_dtype).to(to_device))
   elif src.is_tiny and dest.is_cpu:
     dest.resize_(src.numel()).resize_(src.shape)
@@ -611,7 +615,10 @@ tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
   "aten.fill_.Tensor": lambda self, value: Tensor.full(self.shape, value.reshape(()).item(), device=self.device, dtype=self.dtype),
   "aten.flip": Tensor.flip,
   "aten.scatter_reduce.two": Tensor.scatter_reduce,
-  "aten.squeeze_.dim": lambda self, dim: self.replace(self.squeeze(dim), allow_shape_mismatch=True), # TODO: inplace view op, here?
+  "aten.squeeze_.dim": Tensor.squeeze,
+  "aten.unsqueeze_": Tensor.unsqueeze,
+  "aten.transpose_": Tensor.transpose,
+  "aten.t_": Tensor.transpose,
   "aten.add.Tensor": lambda input,other,alpha=1: input+alpha*other,
   "aten.linspace": lambda start, stop, steps, dtype=None, **kwargs:
     Tensor.linspace(start, stop, steps, **({"dtype": _from_torch_dtype(dtype)} if dtype is not None else {})),
@@ -655,6 +662,13 @@ inplace_ops = {
   "aten.masked_fill_.Tensor",
 }
 
+inplace_view_ops = {
+  "aten.squeeze_.dim",
+  "aten.unsqueeze_",
+  "aten.transpose_",
+  "aten.t_",
+}
+
 def wrap_fxn(k,f):
   def nf(*args, **kwargs):
     if TORCH_DEBUG:
@@ -675,8 +689,42 @@ def wrap_inplace(k,f):
     return orig
   return nf
 
+def wrap_inplace_view_op(k,f):
+  def nf(*args, **kwargs):
+    orig = args[0]
+    args, kwargs = unwrap_args(args, kwargs)
+    target = args[0]
+    new_view = f(*args, **kwargs)
+    if new_view is target or new_view.uop is target.uop:
+      _update_torch_metadata(orig, target)
+      return orig
+    base = canonical_base(target)
+    op = (f, args[1:], kwargs)
+    if target is base:
+      views = derived_views(base)
+      if views:
+        old_base = Tensor(base.uop, device=base.device)
+        old_base.requires_grad = base.requires_grad
+        old_base._views = getattr(base, "_views", set())
+        for v in views: v._view_base = old_base
+        base._views = set()
+        base._view_base = old_base
+        base._view_ops = [op]
+        old_base._views.add(weakref.ref(base))
+    else:
+      target._view_base = base
+      base._views = getattr(base, "_views", set())
+      base._views.add(weakref.ref(target))
+      target._view_ops = _get_view_ops(target) + [op]
+    target.uop = new_view.uop
+    _update_torch_metadata(orig, target)
+    return orig
+  return nf
+
 for k,v in tiny_backend.items():
-  wrapper = wrap_inplace if k in inplace_ops else wrap_fxn
+  if k in inplace_view_ops: wrapper = wrap_inplace_view_op
+  elif k in inplace_ops: wrapper = wrap_inplace
+  else: wrapper = wrap_fxn
   torch.library.impl(k.replace("aten.", "aten::"), "privateuseone")(wrapper(k,v))
 
 @torch.library.impl("aten::equal", "privateuseone")

extra/torch_backend/test_inplace.py

@@ -67,5 +67,37 @@ class TestTorchBackendInplace(unittest.TestCase):
     d += torch.arange(4)
     np.testing.assert_array_equal(a.cpu(), torch.arange(4).cpu())
 
+  def test_inplace_view_metadata(self):
+    a = torch.arange(6, dtype=torch.float32).reshape(1, 2, 3)
+    ret = a.squeeze_(0)
+    self.assertIs(ret, a)
+    self.assertEqual(a.shape, torch.Size([2, 3]))
+    ret = a.unsqueeze_(1)
+    self.assertIs(ret, a)
+    self.assertEqual(a.shape, torch.Size([2, 1, 3]))
+    ret = a.transpose_(0, 2)
+    self.assertIs(ret, a)
+    self.assertEqual(a.shape, torch.Size([3, 1, 2]))
+
+  def test_t_inplace_metadata(self):
+    a = torch.arange(6, dtype=torch.float32).reshape(2, 3)
+    ret = a.t_()
+    self.assertIs(ret, a)
+    self.assertEqual(a.shape, torch.Size([3, 2]))
+    expected = torch.arange(6, dtype=torch.float32).reshape(2, 3).t()
+    np.testing.assert_array_equal(a.cpu().numpy(), expected.cpu().numpy())
+
+  def test_squeeze_matmul(self):
+    # squeeze_ is used internally by PyTorch for vector-matrix matmul (unsqueeze -> mm -> squeeze_)
+    a = torch.arange(65, dtype=torch.float32)
+    b = torch.arange(65*45, dtype=torch.float32).reshape(65, 45)
+    result = a.matmul(b)
+    self.assertEqual(result.shape, torch.Size([45]))
+    # verify correctness
+    a_cpu = torch.arange(65, dtype=torch.float32, device='cpu')
+    b_cpu = torch.arange(65*45, dtype=torch.float32, device='cpu').reshape(65, 45)
+    expected = a_cpu.matmul(b_cpu)
+    np.testing.assert_allclose(result.cpu().numpy(), expected.numpy(), rtol=1e-4, atol=1e-4)
+
 if __name__ == "__main__":
   unittest.main()

extra/torch_backend/wrapped_tensor.cpp

@@ -131,7 +131,15 @@ py::object unwrap_tensor(const at::Tensor &tensor) {
   return py::reinterpret_borrow<py::object>(tiny->ptr(getPyInterpreter()));
 }
 
+void update_metadata(const at::Tensor &tensor, const std::vector<int64_t> &sizes,
+                     const std::vector<int64_t> &strides, int64_t storage_offset) {
+  auto* impl = tensor.unsafeGetTensorImpl();
+  impl->set_allow_tensor_metadata_change(true);
+  impl->set_sizes_and_strides(sizes, strides, storage_offset);
+}
+
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   m.def("wrap", &wrap_tensor);
   m.def("unwrap", &unwrap_tensor);
+  m.def("update_metadata", &update_metadata);
 }

extra/viz/README

@@ -0,0 +1,18 @@
+A command line tool for exploring the VIZ trace.
+
+After running with VIZ=-1, use `PYTHONPATH=. extra/viz/cli.py` to explore the saved trace files.
+
+## Inspect runtime profiling
+
+Use `PYTHONPATH=. extra/viz/cli.py --profile` to list all traced devices.
+
+List top slowest kernels on a device: `--profile --device "AMD"`
+List samples of a kernel on a device: `--profile --device "AMD" --kernel E_3`
+
+## Inspect codegen and PatternMatcher
+
+Use `PYTHONPATH=. extra/viz/cli.py --rewrites` to list all traced kernels.
+
+List all codegen steps for a kernel: `--rewrites --kernel E_3`
+Get source code: `--rewrites --kernel E_3 --select "View Program"`
+Inspect a graph rewrite: `--rewrites --kernel E_3 --select "initial symbolic"`

extra/viz/cli.py

@@ -3,7 +3,8 @@ import argparse, pathlib
 from typing import Iterator
 from tinygrad.viz import serve as viz
 from tinygrad.uop.ops import RewriteTrace
-from tinygrad.helpers import temp, ansistrip, colored
+from tinygrad.helpers import temp, ansistrip, colored, time_to_str, ansilen
+from test.null.test_viz import load_profile
 
 def optional_eq(val:dict, arg:str|None) -> bool: return arg is None or ansistrip(val["name"]) == arg
 
@@ -25,13 +26,61 @@ def print_data(data:dict) -> None:
 
 if __name__ == "__main__":
   parser = argparse.ArgumentParser()
-  parser.add_argument('--kernel', type=str, default=None, metavar="NAME", help='Select a kernel by name (optional name, default: only list names)')
-  parser.add_argument('--select', type=str, default=None, metavar="NAME",
-                      help='Select an item within the chosen kernel (optional name, default: only list names)')
+  g_mode = parser.add_argument_group("mode")
+  g_mode.add_argument("--profile", action="store_true", help="View profile trace")
+  g_mode.add_argument("--rewrites", action="store_true", help="View rewrites trace")
+  g_profile = parser.add_argument_group("profile options")
+  g_profile.add_argument("--device", type=str, default=None, metavar="NAME", help="Select a device (optional name, default: only list names)")
+  g_profile.add_argument("--top", type=int, default=10, metavar="N", help="Number of top kernels to show (-1 for all, default: 10)")
+  g_rewrites = parser.add_argument_group("rewrites options")
+  g_rewrites.add_argument("--select", type=str, default=None, metavar="NAME",
+                          help="Select an item within the chosen kernel (optional name, default: only list names)")
+  g_common = parser.add_argument_group("common options")
+  g_common.add_argument("--kernel", type=str, default=None, metavar="NAME", help="Select a kernel by name (optional name, default: only list names)")
+  parser.add_argument("--profile-path", type=pathlib.Path, metavar="PATH", help="Path to profile (optional file, default: latest profile)",
+                        default=pathlib.Path(temp("profile.pkl", append_user=True)))
+  parser.add_argument("--rewrites-path", type=pathlib.Path, metavar="PATH", help="Path to rewrites (optional file, default: latest rewrites)",
+                        default=pathlib.Path(temp("rewrites.pkl", append_user=True)))
   args = parser.parse_args()
+  if not args.profile and not args.rewrites:
+    parser.print_help()
+    exit(0)
 
-  viz.trace = viz.load_pickle(pathlib.Path(temp("rewrites.pkl", append_user=True)), default=RewriteTrace([], [], {}))
+  viz.trace = viz.load_pickle(args.rewrites_path, default=RewriteTrace([], [], {}))
   viz.ctxs = viz.get_rewrites(viz.trace)
+
+  if args.profile:
+    from tabulate import tabulate
+    profile = load_profile(viz.load_pickle(args.profile_path, default=[]))
+    agg, total, n = {}, 0, 0
+    if args.device is None: print("Select a device:")
+    for k,v in profile["layout"].items():
+      if not optional_eq({"name":k}, args.device): continue
+      print(f"  {k}")
+      if args.device is None: continue
+      for e in v.get("events", []):
+        et = e["dur"]*1e-6
+        if args.kernel is not None:
+          if ansistrip(e["name"]) == args.kernel and n < 10:
+            ptm = colored(time_to_str(et, w=9), "yellow" if et > 0.01 else None) if et is not None else ""
+            name = e["name"]+(" " * (46 - ansilen(e["name"])))
+            print(f"{name} {ptm}/{(et or 0)*1e3:9.2f}ms  "+e['fmt'].replace('\n', ' | ')+"  ")
+            n += 1
+        else:
+          a = agg.setdefault(e["name"], [0.0, 0])
+          a[0] += et
+          a[1] += 1
+          total += et
+    if agg and total > 0:
+      items = sorted(agg.items(), key=lambda kv:kv[1][0], reverse=True)
+      sel = items if args.top == -1 else items[:args.top]
+      table = [[name, time_to_str(t, w=9), c, f"{(t/total*100.0):.2f}%"] for name,(t,c) in sel]
+      if args.top != -1 and (other:=items[len(sel):]):
+        other_t = total-sum(t for _, (t, _) in sel)
+        table.append([f"Other ({len(other)} unique)", time_to_str(other_t, w=9), sum(c for _,(_,c) in other), f"{other_t/total*100.0:.2f}%"])
+      print(tabulate(table, headers=["name", "total", "count", "pct"], tablefmt="github"))
+    exit(0)
+
   for k in viz.ctxs:
     if not optional_eq(k, args.kernel): continue
     print(k["name"])

pyproject.toml

@@ -67,7 +67,7 @@ testing_minimal = [
   "pytest-timeout",
   "pytest-split",
   "hypothesis>=6.148.9",
-  "z3-solver",
+  "z3-solver<4.15.4",  # 4.15.4 has a segfault when creating many z3.Context()
 ]
 testing_unit = ["tinygrad[testing_minimal]", "tqdm", "safetensors", "tabulate", "openai", "ggml-python"]
 testing = [

test/device/test_hcq.py

@@ -119,7 +119,7 @@ class TestHCQ(unittest.TestCase):
     TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
 
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 1.0, f"got val {val}"
 
   def test_exec_2_kernels_100_times(self):
@@ -135,7 +135,7 @@ class TestHCQ(unittest.TestCase):
       q.submit(TestHCQ.d0, {virt_val.expr: TestHCQ.d0.timeline_value})
       TestHCQ.d0.timeline_value += 1
 
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 200.0, f"got val {val}"
 
   @unittest.skipIf(Device.DEFAULT in {"CPU"}, "No globals/locals on LLVM/CPU")
@@ -151,9 +151,9 @@ class TestHCQ(unittest.TestCase):
     TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
 
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 1.0, f"got val {val}"
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
     assert val == 0.0, f"got val {val}, should not be updated"
 
   @unittest.skipIf(Device.DEFAULT in {"CPU"}, "No globals/locals on LLVM/CPU")
@@ -186,7 +186,7 @@ class TestHCQ(unittest.TestCase):
           TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
           TestHCQ.d0.timeline_value += 1
 
-          res_sum = sum(x for x in zt.as_buffer().cast("I"))
+          res_sum = sum(x for x in zt.as_memoryview().cast("I"))
           assert x * y * z == res_sum, f"want {x * y * z}, got {res_sum}"
 
   # Test copy
@@ -200,7 +200,7 @@ class TestHCQ(unittest.TestCase):
     TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
 
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
     assert val == 1.0, f"got val {val}"
 
   def test_copy_long(self):
@@ -218,7 +218,7 @@ class TestHCQ(unittest.TestCase):
     TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
 
-    mv_buf1 = buf1.as_buffer().cast('Q')
+    mv_buf1 = buf1.as_memoryview().cast('Q')
     assert libc.memcmp(mv_address(mv_buf1), buf2._buf.va_addr, sz) == 0
 
   @slow
@@ -242,7 +242,7 @@ class TestHCQ(unittest.TestCase):
       TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
 
-      mv_buf1 = buf1.as_buffer()
+      mv_buf1 = buf1.as_memoryview()
       assert libc.memcmp(mv_address(mv_buf1), buf2._buf.va_addr, sz) == 0
 
   def test_update_copy(self):
@@ -260,7 +260,7 @@ class TestHCQ(unittest.TestCase):
     TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
 
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
     assert val == 1.0, f"got val {val}"
 
   def test_update_copy_long(self):
@@ -283,7 +283,7 @@ class TestHCQ(unittest.TestCase):
     TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
 
-    mv_buf1 = buf1.as_buffer().cast('Q')
+    mv_buf1 = buf1.as_memoryview().cast('Q')
     for i in range(sz//8): assert mv_buf1[i] == 0x0101010101010101, f"offset {i*8} differs, not all copied, got {hex(mv_buf1[i])}"
 
   # Test bind api
@@ -421,7 +421,7 @@ class TestHCQ(unittest.TestCase):
       TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
 
-      assert buf1.as_buffer()[0] == i
+      assert buf1.as_memoryview()[0] == i
 
   def test_small_copies_from_host_buf_intercopy(self):
     if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -440,7 +440,7 @@ class TestHCQ(unittest.TestCase):
       TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
 
-      assert buf2.as_buffer()[0] == i
+      assert buf2.as_memoryview()[0] == i
 
   def test_small_copies_from_host_buf_transfer(self):
     if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -463,7 +463,7 @@ class TestHCQ(unittest.TestCase):
       TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
 
-      assert buf2.as_buffer()[0] == i
+      assert buf2.as_memoryview()[0] == i
 
   def test_memory_barrier(self):
     a = Tensor([0, 1], device=Device.DEFAULT, dtype=dtypes.int8).realize()
@@ -486,7 +486,7 @@ class TestHCQ(unittest.TestCase):
       TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
 
-      assert buf1.as_buffer()[0] == (i + 1), f"has {buf1.as_buffer()[0]}, need {i + 1}"
+      assert buf1.as_memoryview()[0] == (i + 1), f"has {buf1.as_memoryview()[0]}, need {i + 1}"
 
   def test_memory_barrier_before_copy(self):
     if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -511,7 +511,7 @@ class TestHCQ(unittest.TestCase):
       TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
 
-      assert buf2.as_buffer()[0] == i
+      assert buf2.as_memoryview()[0] == i
 
   def test_map_cpu_buffer_to_device(self):
     if Device[Device.DEFAULT].hw_copy_queue_t is None: self.skipTest("skip device without copy queue")

test/device/test_metal.py

@@ -1,5 +1,5 @@
 import unittest
-from tinygrad.device import CompileError, Device, Compiler
+from tinygrad.device import CompileError, Device
 if Device.DEFAULT=="METAL":
   from tinygrad.runtime.ops_metal import MetalDevice, MetalCompiler, MetalProgram
 @unittest.skipIf(Device.DEFAULT!="METAL", "Metal support required")
@@ -48,28 +48,4 @@ kernel void r_5(device int* data0, const device int* data1, uint3 gid [[threadgr
 """)
     with self.assertRaises(RuntimeError):
       compiled = compiled[:40] # corrupt the compiled program
-      MetalProgram(device, "r_5", compiled)
-
-  def test_program_w_empty_compiler(self):
-    device = MetalDevice("metal")
-    compiler = Compiler(device)
-    compiled = compiler.compile("""
-#include <metal_stdlib>
-kernel void r_5(device int* data0, const device int* data1, uint3 gid [[threadgroup_position_in_grid]], uint3 lid [[thread_position_in_threadgroup]]){
-  data0[0] = 0;
-}
-""")
-    MetalProgram(device, "r_5", compiled)
-
-  def test_bad_program_w_empty_compiler(self):
-    device = MetalDevice("metal")
-    compiler = Compiler(device)
-    # this does not raise
-    compiled = compiler.compile("""
-#include <metal_stdlib>
-kernel void r_5(device int* data0, const device int* data1, uint3 gid [[threadgroup_position_in_grid]], uint3 lid [[thread_position_in_threadgroup]]){
-  invalid codes;
-}
-""")
-    with self.assertRaises(RuntimeError):
       MetalProgram(device, "r_5", compiled)

test/device/test_ocl.py

@@ -1,9 +1,19 @@
 import unittest
+from unittest.mock import patch
 from tinygrad import Device
 from tinygrad.device import Buffer
 from tinygrad.dtype import dtypes
 from tinygrad.runtime.ops_cl import CLDevice, CLAllocator, CLCompiler, CLProgram
 
+@unittest.skipUnless(Device.DEFAULT == "CL", "Runs only on OpenCL")
+class TestCLCompileCache(unittest.TestCase):
+  def test_compile_cached(self):
+    device = Device[Device.DEFAULT]
+    src = "__kernel void cached_test(__global int* a) { a[0] = 1; }"
+    CLProgram(device, name="cached_test", lib=src.encode())
+    with patch.object(CLCompiler, 'compile', side_effect=RuntimeError("compile should not be called on cache hit")):
+      CLProgram(device, name="cached_test", lib=src.encode())
+
 @unittest.skipUnless(Device.DEFAULT == "CL", "Runs only on OpenCL")
 class TestCLError(unittest.TestCase):
   @unittest.skip("allocates tons of memory")
@@ -17,7 +27,7 @@ class TestCLError(unittest.TestCase):
   def test_invalid_kernel_name(self):
     device = Device[Device.DEFAULT]
     with self.assertRaises(RuntimeError) as err:
-      CLProgram(device, name="", lib=CLCompiler(device, "test").compile("__kernel void test(__global int* a) { a[0] = 1; }"))
+      CLProgram(device, name="", lib="__kernel void test(__global int* a) { a[0] = 1; }".encode())
     assert str(err.exception) == "OpenCL Error -46: CL_INVALID_KERNEL_NAME"
 
   def test_unaligned_copy(self):

test/external/external_fuzz_qcom_cpu_cache.py

@@ -0,0 +1,11 @@
+#!/usr/bin/env python3
+from tinygrad.tensor import Tensor
+import numpy as np
+
+while True:
+  arr = np.ones(1000000, dtype=np.uint8)
+  print(f"numpy: {(arr + 1)[:10]}")
+
+  ptr = arr.ctypes.data
+  tensor = Tensor.from_blob(ptr, arr.shape, dtype='uint8', device='QCOM').realize() + 1
+  print(f"from_blob: {tensor.numpy()[:10]}")

test/external/external_test_amd.py

@@ -20,7 +20,7 @@ class TestAMD(unittest.TestCase):
                               global_size=TestAMD.d0_runner.global_size, local_size=TestAMD.d0_runner.local_size)
       TestAMD.d0_runner.clprg(TestAMD.a.uop.buffer._buf, TestAMD.b.uop.buffer._buf,
                               global_size=TestAMD.d0_runner.global_size, local_size=TestAMD.d0_runner.local_size)
-    val = TestAMD.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestAMD.a.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 4000.0, f"got val {val}"
 
 if __name__ == "__main__":

test/external/external_test_hcq.py

@@ -65,7 +65,7 @@ class TestHCQ(unittest.TestCase):
     q.submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 2000.0, f"got val {val}"
 
   def test_run_1000_times(self):
@@ -81,7 +81,7 @@ class TestHCQ(unittest.TestCase):
       TestHCQ.compute_queue().signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
       TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 2000.0, f"got val {val}"
 
   def test_run_to_3(self):
@@ -95,7 +95,7 @@ class TestHCQ(unittest.TestCase):
     q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 3.0, f"got val {val}"
 
   def test_update_exec(self):
@@ -106,9 +106,9 @@ class TestHCQ(unittest.TestCase):
     q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 1.0, f"got val {val}"
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
     assert val == 0.0, f"got val {val}, should not be updated"
 
   @unittest.skipUnless(Device.DEFAULT == "NV", "Only NV supports bind")
@@ -126,7 +126,7 @@ class TestHCQ(unittest.TestCase):
       TestHCQ.compute_queue().signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
       TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 2000.0, f"got val {val}"
 
   @unittest.skipUnless(Device.DEFAULT == "NV", "Only NV supports bind")
@@ -141,9 +141,9 @@ class TestHCQ(unittest.TestCase):
     q.submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 1.0, f"got val {val}"
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
     assert val == 0.0, f"got val {val}, should not be updated"
 
   @unittest.skipIf(CI, "Can't handle async update on CPU")
@@ -174,7 +174,7 @@ class TestHCQ(unittest.TestCase):
     q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 1.0, f"got val {val}"
 
   def test_submit_empty_queues(self):
@@ -206,7 +206,7 @@ class TestHCQ(unittest.TestCase):
     q.submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 1.0, f"got val {val}"
 
   def test_copy_1000_times(self):
@@ -221,7 +221,7 @@ class TestHCQ(unittest.TestCase):
     # confirm the signal didn't exceed the put value
     with self.assertRaises(RuntimeError):
       TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value + 1, timeout=50)
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
     assert val == 0.0, f"got val {val}"
 
   def test_copy(self):
@@ -231,7 +231,7 @@ class TestHCQ(unittest.TestCase):
     q.submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
     assert val == 1.0, f"got val {val}"
 
   @unittest.skipUnless(Device.DEFAULT == "NV", "Only NV supports bind")
@@ -248,7 +248,7 @@ class TestHCQ(unittest.TestCase):
     # confirm the signal didn't exceed the put value
     with self.assertRaises(RuntimeError):
       TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value + 1, timeout=50)
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
     assert val == 0.0, f"got val {val}"
 
   def test_copy_bandwidth(self):
@@ -288,7 +288,7 @@ class TestHCQ(unittest.TestCase):
     q.submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
     assert val == 1.0, f"got val {val}"
 
   def test_cross_device_signal(self):
@@ -319,7 +319,7 @@ class TestHCQ(unittest.TestCase):
       q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
       TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1
-      val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+      val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
       assert val == 1.0, f"got val {val}"
 
 if __name__ == "__main__":

test/external/external_test_schedule_scaling.py

@@ -0,0 +1,134 @@
+import unittest, time
+from tinygrad import Tensor
+
+class TestScheduleScaling(unittest.TestCase):
+  """Test that .schedule() scales linearly with graph size (no O(n^2) behavior)."""
+
+  def _assert_linear(self, fn, n_small=200, n_large=1000):
+    """Assert schedule time scales at most ~linearly: time(n_large)/time(n_small) should be close to n_large/n_small."""
+    fn(n_small).schedule()  # warmup
+    t_small = min(self._time_schedule(fn, n) for n in [n_small]*3)
+    t_large = min(self._time_schedule(fn, n) for n in [n_large]*3)
+    size_ratio = n_large / n_small  # 5.0
+    time_ratio = t_large / t_small
+    # O(n) -> time_ratio ~ 5, O(n^2) -> time_ratio ~ 25. threshold at 10 catches n^2 with margin.
+    self.assertLess(time_ratio / size_ratio, 2.0,
+      f"schedule appears superlinear: n={n_small} {t_small*1e3:.1f}ms, n={n_large} {t_large*1e3:.1f}ms "
+      f"(time grew {time_ratio:.1f}x for {size_ratio:.0f}x size, per-node ratio {time_ratio/size_ratio:.2f})")
+
+  @staticmethod
+  def _time_schedule(fn, n) -> float:
+    st = time.perf_counter()
+    fn(n).schedule()
+    return time.perf_counter() - st
+
+  # *** rangeify: ending_ranges accumulation and consumer merge ***
+
+  # ending_ranges accumulation via sum([], []) and nested scan in run_rangeify.
+  # this creates reduce ops whose ending_ranges lists grow with graph depth, causing O(n^2) list copies.
+  def test_multi_reduce_scaling(self):
+    def multi_reduce(n):
+      x = Tensor.empty(256, 256)
+      for _ in range(n):
+        s = x.sum(axis=-1, keepdim=True)
+        x = x + s + s
+      return x
+    self._assert_linear(multi_reduce)
+
+  # reduce+elementwise chain stresses ending_ranges propagation and post-rangeify rewrites
+  def test_wide_reduce_scaling(self):
+    def wide_reduce(n):
+      x = Tensor.empty(256, 256)
+      for _ in range(n):
+        x = x + x.sum(axis=-1, keepdim=True)
+      return x
+    self._assert_linear(wide_reduce)
+
+  # expand ops inject into ending_ranges via the EXPAND path in run_rangeify
+  def test_expand_reduce_scaling(self):
+    def expand_reduce(n):
+      x = Tensor.empty(256, 1)
+      for _ in range(n):
+        y = x.expand(256, 256)
+        x = (y + y).sum(axis=-1, keepdim=True)
+      return x
+    self._assert_linear(expand_reduce)
+
+  # *** graph_rewrite: multi-consumer DAG patterns ***
+
+  # multi-consumer diamond pattern (fan-out/fan-in) stresses consumer_rngs merge in run_rangeify
+  def test_diamond_scaling(self):
+    def diamond(n):
+      x = Tensor.empty(256, 256)
+      for _ in range(n):
+        a = x + 1
+        b = x + 2
+        x = a + b
+      return x
+    self._assert_linear(diamond)
+
+  # elementwise chain baseline — should be trivially O(n)
+  def test_chain_scaling(self):
+    def chain(n):
+      x = Tensor.empty(256, 256)
+      for _ in range(n): x = x + 1
+      return x
+    self._assert_linear(chain)
+
+  # softmax has multi-consumer structure (x used for max, exp, and sum), stresses graph_rewrite on DAGs
+  def test_softmax_scaling(self):
+    def softmax_chain(n):
+      x = Tensor.empty(64, 256)
+      for _ in range(n): x = x.softmax(axis=-1)
+      return x
+    self._assert_linear(softmax_chain)
+
+  # *** post-rangeify: symbolic rewrites, kernel splitting ***
+
+  # matmul chain stresses symbolic+reduce_collapse and split_store
+  def test_matmul_scaling(self):
+    def matmul_chain(n):
+      xs = [Tensor.empty(32, 32) for _ in range(n + 1)]
+      result = xs[0]
+      for i in range(n): result = result @ xs[i + 1]
+      return result
+    self._assert_linear(matmul_chain)
+
+  # contiguous chain stresses remove_bufferize callbacks (toposort per BUFFERIZE node)
+  def test_contiguous_scaling(self):
+    def contiguous_chain(n):
+      x = Tensor.empty(256, 256)
+      for _ in range(n): x = (x + 1).contiguous()
+      return x
+    self._assert_linear(contiguous_chain)
+
+  # *** schedule: AFTER handling, assign ***
+
+  # assign chain stresses AFTER cycle detection (toposort inside toposort loop in get_rangeify_map)
+  def test_assign_scaling(self):
+    def assign_chain(n):
+      x = Tensor.empty(256, 256).realize()
+      for _ in range(n): x.assign(x + 1)
+      return x
+    self._assert_linear(assign_chain)
+
+  # layernorm has multi-consumer reduces (mean reused in variance), stresses consumer_rngs merge and symbolic rewrites
+  def test_layernorm_scaling(self):
+    def layernorm_chain(n):
+      x = Tensor.empty(64, 256)
+      for _ in range(n):
+        mean = x.mean(axis=-1, keepdim=True)
+        var = ((x - mean) ** 2).mean(axis=-1, keepdim=True)
+        x = (x - mean) / (var + 1e-5).sqrt()
+      return x
+    self._assert_linear(layernorm_chain)
+
+  # concat chain stresses MSTACK/MSELECT handling and wide SINK construction
+  def test_concat_scaling(self):
+    def concat_chain(n):
+      parts = [Tensor.empty(4, 256) + i for i in range(n)]
+      return parts[0].cat(*parts[1:])
+    self._assert_linear(concat_chain)
+
+if __name__ == '__main__':
+  unittest.main(verbosity=2)

test/external/fuzz_graph.py

@@ -30,7 +30,7 @@ def alloc_rawbuffer(device, fill=False):
   if fill:
     with Context(DEBUG=0):
       data = np.random.randint(-10000, 10000, size=rawbuf.size, dtype=_to_np_dtype(rawbuf.dtype))
-      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_buffer())
+      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_memoryview())
   return rawbuf
 
 def gen_kernel_ji(device, deps):
@@ -93,7 +93,7 @@ def run_jit(jis, all_buffers, input_buffers, var_vals):
 
   with Context(DEBUG=0):
     res_buffers = []
-    for rawbuf in all_buffers: res_buffers.append(rawbuf.as_buffer())
+    for rawbuf in all_buffers: res_buffers.append(rawbuf.as_memoryview())
     return res_buffers
 
 def fuzz_graph(jis, all_buffers, input_buffers):

test/external/fuzz_symbolic.py

@@ -1,12 +1,16 @@
-import random, operator
+# NOTE: z3-solver 4.15.4 segfaults (exit code 139) when creating many z3.Context() with complex expressions.
+# Reproduces consistently with seed=74 around iteration 1767. Versions <=4.15.3 are fine.
+# Workaround: reuse a single z3.Context, or pin z3-solver<4.15.4 (see pyproject.toml).
+# To repro: pip install z3-solver==4.15.4.0 && python test/external/fuzz_symbolic.py 74
+import random, operator, sys
 import z3
 from tinygrad import Variable, dtypes
 from tinygrad.uop.ops import UOp
 from tinygrad.uop.validate import uops_to_z3
 from tinygrad.helpers import DEBUG, Context
 
-seed = random.randint(0, 100)
-print(f"Seed: {seed}")
+seed = int(sys.argv[1]) if len(sys.argv) > 1 else random.randint(0, 100)
+print(f"Seed: {seed}", flush=True)
 random.seed(seed)
 
 unary_ops = [lambda a:a+random.randint(-4, 4), lambda a: a*random.randint(-4, 4),

test/external/fuzz_symbolic_symbolic_div.py

@@ -1,11 +1,11 @@
-import random
+import random, sys
 import z3
 from tinygrad.uop.ops import UOp, Ops
 from tinygrad.uop.validate import uops_to_z3
 from tinygrad.helpers import DEBUG, Context, colored
 
-seed = random.randint(0, 100)
-print(f"Seed: {seed}")
+seed = int(sys.argv[1]) if len(sys.argv) > 1 else random.randint(0, 100)
+print(f"Seed: {seed}", flush=True)
 random.seed(seed)
 
 def get_random_term(ranges, factors):

test/external/process_replay/process_replay.py

@@ -47,13 +47,15 @@ def replay_get_rangeify_map(ret:dict[UOp, UOp], big_sink:UOp) -> tuple[str, str,
   UOp.unique_num = itertools.count(max([u.arg for u in big_sink.toposort() if u.op is Ops.UNIQUE], default=0)+1)
   new_sink = big_sink.substitute(get_rangeify_map(big_sink))
   def to_str(ret:UOp) -> str:
-    asts = [repr(u.arg.ast) for u in ret.toposort() if u.op is Ops.KERNEL]
+    asts = [repr(u.arg.ast) for u in ret.toposort() if u.op is Ops.CALL]
     return "\n".join([f"{len(asts)} kernels", *asts])
   return to_str(new_sink), to_str(big_sink.substitute(ret)), (big_sink,)
 
 def replay_get_program(p:ProgramSpec, ast:UOp, renderer:Renderer, opts:list[Opt]|None=None) -> tuple[str, str, tuple[Any, ...]]:
   # the ast.arg is non None if we are inside of search.py
-  sink_arg = ast.arg or KernelInfo(opts_to_apply=tuple(opts) if opts is not None else p.applied_opts if BEAM>=1 else None)
+  sink_arg = ast.arg or KernelInfo()
+  if opts is not None: sink_arg = replace(sink_arg, opts_to_apply=tuple(opts))
+  elif BEAM >= 1 and sink_arg.opts_to_apply is None: sink_arg = replace(sink_arg, opts_to_apply=p.applied_opts)
   input_ast = ast if ast.op is Ops.PROGRAM else ast.replace(arg=replace(sink_arg, name=p.name))
   p2 = get_program(input_ast, renderer=renderer)
   def to_str(ret:ProgramSpec) -> str:

test/helpers.py

@@ -62,7 +62,7 @@ def eval_uop(uop:UOp, inputs:list[tuple[DType, list[Any]]]|None=None):
     bufs.append(buf:=allocator.alloc(len(data) * buf_dt.itemsize))
     allocator._copyin(buf, memoryview(struct.pack(str(len(data)) + (buf_dt.fmt or ""), *data)))
   g = UOp(Ops.PARAM, uop.dtype.ptr(), arg=0, src=())
-  prg = get_program(UOp.store(g.index(UOp.const(dtypes.int, 0)), uop).sink(), PythonRenderer())
+  prg = get_program(UOp.store(g.index(UOp.const(dtypes.int, 0)), uop).sink(arg=KernelInfo()), PythonRenderer())
   prog = PythonProgram("run", PythonCompiler().compile(prg.src))
   prog(out_buf:=allocator.alloc(uop.dtype.itemsize), *bufs)
   return out_buf.cast(uop.dtype.fmt or "").tolist()[0]

test/mockgpu/nv/nvdriver.py

@@ -130,15 +130,18 @@ class NVDriver(VirtDriver):
       struct.hObjectNew = self._alloc_handle()
       self.object_by_handle[struct.hObjectNew] = NVContextShare(self.object_by_handle[struct.hObjectParent])
     elif struct.hClass == nv_gpu.AMPERE_CHANNEL_GPFIFO_A:
-      assert struct.hObjectParent in self.object_by_handle and isinstance(self.object_by_handle[struct.hObjectParent], NVChannelGroup)
+      parent = self.object_by_handle.get(struct.hObjectParent)
+      assert parent is not None and isinstance(parent, (NVChannelGroup, NVGPU))
       struct.hObjectNew = self._alloc_handle()
       params = nv_gpu.NV_CHANNELGPFIFO_ALLOCATION_PARAMETERS.from_address(params_ptr)
-      gpu = self.object_by_handle[struct.hObjectParent].device
+      gpu = parent.device if isinstance(parent, NVChannelGroup) else parent
       gpfifo_token = gpu.add_gpfifo(params.gpFifoOffset, params.gpFifoEntries)
       self.object_by_handle[struct.hObjectNew] = NVGPFIFO(gpu, gpfifo_token)
-    elif struct.hClass == nv_gpu.AMPERE_DMA_COPY_B or struct.hClass == nv_gpu.ADA_COMPUTE_A:
+    elif struct.hClass in (nv_gpu.AMPERE_DMA_COPY_B, nv_gpu.ADA_COMPUTE_A, nv_gpu.NVC9B0_VIDEO_DECODER, nv_gpu.NVCFB0_VIDEO_DECODER):
       assert struct.hObjectParent in self.object_by_handle and isinstance(self.object_by_handle[struct.hObjectParent], NVGPFIFO)
       struct.hObjectNew = self._alloc_handle()
+      gpfifo = self.object_by_handle[struct.hObjectParent]
+      gpfifo.device.queues[gpfifo.token].bound_engines.add(struct.hClass)
     elif struct.hClass == nv_gpu.GT200_DEBUGGER:
       struct.hObjectNew = self._alloc_handle()
     elif struct.hClass == nv_gpu.MAXWELL_PROFILER_DEVICE:
@@ -194,7 +197,7 @@ class NVDriver(VirtDriver):
       params = nv_gpu.NVC36F_CTRL_CMD_GPFIFO_GET_WORK_SUBMIT_TOKEN_PARAMS.from_address(params_ptr)
       gpu_fifo = self.object_by_handle[struct.hObject]
       params.workSubmitToken = gpu_fifo.token
-    elif struct.cmd == nv_gpu.NVA06C_CTRL_CMD_GPFIFO_SCHEDULE: pass
+    elif struct.cmd in (nv_gpu.NVA06C_CTRL_CMD_GPFIFO_SCHEDULE, nv_gpu.NVA06F_CTRL_CMD_BIND, nv_gpu.NVA06F_CTRL_CMD_GPFIFO_SCHEDULE): pass
     elif struct.cmd == nv_gpu.NV2080_CTRL_CMD_PERF_BOOST: pass
     elif struct.cmd == nv_gpu.NV2080_CTRL_CMD_FB_FLUSH_GPU_CACHE: pass
     elif struct.cmd == nv_gpu.NV83DE_CTRL_CMD_DEBUG_READ_ALL_SM_ERROR_STATES:

test/mockgpu/nv/nvgpu.py

@@ -26,6 +26,7 @@ class GPFIFO:
     self.gpfifo = to_mv(self.base, self.entries_cnt * 8).cast("Q")
     self.ctrl = nv_gpu.AmpereAControlGPFifo.from_address(self.base + self.entries_cnt * 8)
     self.state = {}
+    self.bound_engines: set[int] = set()
 
     # Buf exec state
     self.buf = None
@@ -115,9 +116,11 @@ class GPFIFO:
   def execute_cmd(self, cmd) -> SchedResult:
     if cmd == nv_gpu.NVC56F_SEM_EXECUTE: return self._exec_signal()
     elif cmd == nv_gpu.NVC6C0_LAUNCH_DMA: return self._exec_nvc6c0_dma()
-    elif cmd == nv_gpu.NVC6B5_LAUNCH_DMA: return self._exec_nvc6b5_dma()
+    elif cmd == nv_gpu.NVC6B5_LAUNCH_DMA: # NOTE: NVC6B5_LAUNCH_DMA == NVC9B0_EXECUTE == 0x300
+      return self._exec_vid_decode() if self.bound_engines & {nv_gpu.NVC9B0_VIDEO_DECODER, nv_gpu.NVCFB0_VIDEO_DECODER} else self._exec_nvc6b5_dma()
     elif cmd == nv_gpu.NVC6C0_SEND_SIGNALING_PCAS2_B: return self._exec_pcas2()
     elif cmd == 0x0320: return self._exec_load_inline_qmd() # NVC6C0_LOAD_INLINE_QMD_DATA
+    elif cmd == nv_gpu.NVC9B0_SEMAPHORE_D: return self._exec_vid_semaphore()
     else: self.state[cmd] = self._next_dword() # just state update
     return SchedResult.CONT
 
@@ -136,6 +139,27 @@ class GPFIFO:
     else: raise RuntimeError(f"Unsupported type={typ} in exec wait/signal")
     return SchedResult.CONT
 
+  def _exec_vid_decode(self) -> SchedResult:
+    self._next_dword() # consume execute flags
+    # validate that all required decode state was set up correctly
+    assert self._state(nv_gpu.NVC9B0_SET_APPLICATION_ID) == nv_gpu.NVC9B0_SET_APPLICATION_ID_ID_HEVC
+    pic_desc_addr = self._state(nv_gpu.NVC9B0_SET_DRV_PIC_SETUP_OFFSET) << 8
+    pic = nv_gpu.nvdec_hevc_pic_s.from_address(pic_desc_addr)
+    assert pic.stream_len > 0 and pic.pic_width_in_luma_samples > 0 and pic.pic_height_in_luma_samples > 0
+    assert self._state(nv_gpu.NVC9B0_SET_IN_BUF_BASE_OFFSET) != 0
+    assert self._state(nv_gpu.NVC9B0_SET_COLOC_DATA_OFFSET) != 0
+    assert self._state(nv_gpu.NVC9B0_SET_NVDEC_STATUS_OFFSET) != 0
+    assert self._state(nv_gpu.NVC9B0_HEVC_SET_FILTER_BUFFER_OFFSET) != 0
+    return SchedResult.CONT
+
+  def _exec_vid_semaphore(self) -> SchedResult:
+    signal = self._state64(nv_gpu.NVC9B0_SEMAPHORE_A)
+    val = self._state(nv_gpu.NVC9B0_SEMAPHORE_C)
+    self._next_dword() # flags
+    to_mv(signal, 8).cast('Q')[0] = val
+    to_mv(signal + 8, 8).cast('Q')[0] = int(time.perf_counter() * 1e9)
+    return SchedResult.CONT
+
   def _exec_load_inline_qmd(self):
     qmd_addr = self._state64(nv_gpu.NVC6C0_SET_INLINE_QMD_ADDRESS_A) << 8
     assert qmd_addr != 0x0, f"invalid qmd address {qmd_addr}"

test/models/test_efficientnet.py

@@ -72,6 +72,7 @@ class TestEfficientNet(unittest.TestCase):
     self.assertEqual(_LABELS[labels[0]], "hen")
     self.assertEqual(_LABELS[labels[1]], "sports car, sport car")
 
+@unittest.skip("these pretrained models are no longer available")
 class TestViT(unittest.TestCase):
   @classmethod
   def setUpClass(cls):

test/models/test_onnx.py

@@ -73,7 +73,7 @@ class TestHuggingFaceOnnxModels(unittest.TestCase):
     onnx_model_path = snapshot_download_with_retry(
       repo_id=repo_id,
       allow_patterns=["*.onnx", "*.onnx_data"],
-      cache_dir=str(DOWNLOADS_DIR)
+      local_dir=DOWNLOADS_DIR / repo_id
     )
     onnx_model_path = onnx_model_path / model_file
     file_size = onnx_model_path.stat().st_size

test/null/test_attention.py

@@ -0,0 +1,39 @@
+import unittest
+from tinygrad import Tensor, dtypes, TinyJit, UOp
+from tinygrad.apps.llm import apply_rope as apply_rope_new, precompute_freqs_cis
+
+def apply_rope(x:Tensor, start_pos:int):
+  B, H, T, Hd = x.shape
+  precompute_freqs_cis.cache_clear()
+  freqs_cis = precompute_freqs_cis(Hd, start_pos+T)[start_pos:start_pos+T]
+  return apply_rope_new(x, freqs_cis)
+
+class TestAttention(unittest.TestCase):
+  def test_half_qkv_buffers(self):
+    BS, seqlen, dim = 10, 4, 100
+    q = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
+    k = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
+    v = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
+    attn = q.scaled_dot_product_attention(k, v)
+    sched = attn.schedule()
+    # attention has 4 kernels now
+    self.assertEqual(len(sched), 4)
+
+  def test_apply_rope_jit_prune(self):
+    def rope_fn(x_in, pos): return apply_rope(x_in, pos)
+    rope_noprune = TinyJit(rope_fn)
+    rope_prune = TinyJit(rope_fn, prune=True)
+
+    v_pos = UOp.variable("start_pos", 0, 100)
+    for _ in range(3):
+      rope_noprune(Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32), v_pos.bind(1))
+      rope_prune(Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32), v_pos.bind(1))
+    noprune_size = len(rope_noprune.captured.jit_cache)
+    prune_size = len(rope_prune.captured.jit_cache)
+
+    self.assertGreater(noprune_size, prune_size)
+    self.assertGreaterEqual(noprune_size, 2)
+    self.assertEqual(prune_size, 1)
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_const_folding.py

@@ -0,0 +1,149 @@
+import unittest, itertools, math
+from tinygrad import Tensor, dtypes, Context
+from tinygrad.dtype import DType, ConstType
+from tinygrad.uop.ops import Ops, UOp
+from tinygrad.codegen import full_rewrite_to_sink
+import numpy as np
+
+def _check_ast_count(desired_count:int, t:Tensor):
+  # NOTE: this has side effect because everything can be scheduled only once
+  schedule = t.schedule()
+  asts = [s for s in schedule if s.ast.op is Ops.SINK]
+  assert len(asts) == desired_count, f"{len(asts)} != {desired_count}"
+
+class TestUnaryOpsConstFolding(unittest.TestCase):
+  def test_all_consts_ops(self):
+    _check_ast_count(0, Tensor.ones(4).exp())
+    _check_ast_count(0, Tensor.ones(4).sqrt())
+    _check_ast_count(0, Tensor.ones(4) + Tensor.ones(4))
+    _check_ast_count(0, Tensor.ones(4) / Tensor.ones(4))
+
+  def test_cast(self):
+    _check_ast_count(0, Tensor.ones(4).cast(dtypes.int16))
+    _check_ast_count(0, Tensor.full(4, fill_value=-1).cast(dtypes.uint16))
+
+  def test_neg_folding(self):
+    _check_ast_count(0, Tensor([1, 2, 3]).mul(-1).neg())
+    _check_ast_count(0, Tensor([1, 2, 3]).neg().mul(-1))
+    _check_ast_count(0, Tensor([1, 2, 3]).neg().neg())
+
+  def test_neg_realized_no_fold(self):
+    x = Tensor.randn(32, 32)
+    x = x.clip(0, 1).realize()
+    _check_ast_count(1, x.neg())
+
+class TestBinaryOpsConstFolding(unittest.TestCase):
+  def test_add_literal_zero(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) + 0)
+  def test_add_tensor_zero(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) + Tensor.zeros(4))
+  def test_literal_zero_add(self):
+    _check_ast_count(0, 0 + Tensor([1.0, 2, 3, 4]))
+  def test_tensor_zero_add(self):
+    _check_ast_count(0, Tensor.zeros(4) + Tensor([1.0, 2, 3, 4]))
+
+  def test_sub_literal_zero(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) - 0)
+  def test_sub_tensor_zero(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) - Tensor.zeros(4))
+
+  def test_mul_literal_zero(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) * 0)
+  def test_mul_tensor_zero(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) * Tensor.zeros(4))
+  def test_literal_zero_mul(self):
+    _check_ast_count(0, 0 * Tensor([1.0, 2, 3, 4]) * 0)
+  def test_tensor_zero_mul(self):
+    _check_ast_count(0, Tensor.zeros(4) * Tensor([1.0, 2, 3, 4]))
+
+  def test_mul_literal_one(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) * 1)
+  def test_mul_tensor_one(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) * Tensor.ones(4))
+  def test_literal_one_mul(self):
+    _check_ast_count(0, 1 * Tensor([1.0, 2, 3, 4]))
+  def test_tensor_one_mul(self):
+    _check_ast_count(0, Tensor.ones(4) * Tensor([1.0, 2, 3, 4]))
+
+  def test_bool_tensor_mul_bool(self):
+    _check_ast_count(0, Tensor([True, False]) * True)
+    _check_ast_count(0, Tensor([True, False]) * False)
+  def test_bool_mul_bool_tensor(self):
+    _check_ast_count(0, True * Tensor([True, False]))
+    _check_ast_count(0, False * Tensor([True, False]))
+
+  def test_div_literal_one(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) / 1)
+  def test_div_tensor_one(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) / Tensor.ones(4))
+
+  def test_idiv_literal_one(self):
+    _check_ast_count(0, Tensor([1, 2, 3, 4]) // 1)
+  def test_idiv_tensor_one(self):
+    _check_ast_count(0, Tensor([1, 2, 3, 4]) // Tensor.ones(4, dtype=dtypes.int32))
+
+  def test_pow_literal_zero(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) ** 0)
+  def test_pow_tensor_zero(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) ** Tensor.zeros(4))
+
+  def test_pow_literal_one(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) ** 1)
+  def test_pow_tensor_one(self):
+    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) ** Tensor.ones(4))
+  def test_literal_one_pow(self):
+    _check_ast_count(0, 1 ** Tensor([1.0, 2, 3, 4]))
+  def test_tensor_one_pow(self):
+    _check_ast_count(0, Tensor.ones(4) ** Tensor([1.0, 2, 3, 4]))
+
+class TestBitcastConstFolding(unittest.TestCase):
+  def test_scalar_bitcast(self):
+    def t(cases: dict[DType, ConstType]):
+      for (from_dt, from_v), (to_dt, to_v) in itertools.product(cases.items(), cases.items()):
+        if not math.isnan(from_v):
+          r = full_rewrite_to_sink(UOp.const(from_dt, from_v).bitcast(to_dt).sink()).src[0]
+          self.assertEqual(r.op, Ops.CONST, msg:=f"{from_dt} -> {to_dt} ({from_v} -> {to_v})")
+          self.assertEqual(r.dtype, to_dt, msg)
+          np.testing.assert_equal(r.arg, to_v, msg)
+
+    t({dtypes.int8: 0, dtypes.uint8: 0, dtypes.bool: False})
+    t({dtypes.int8: 1, dtypes.uint8: 1, dtypes.bool: True})
+
+    t({dtypes.int8:  -1, dtypes.uint8:  2**8-1})
+    t({dtypes.int16: -1, dtypes.uint16: 2**16-1, dtypes.float16: float('nan')})
+    t({dtypes.int32: -1, dtypes.uint32: 2**32-1, dtypes.float32: float('nan')})
+    t({dtypes.int64: -1, dtypes.uint64: 2**64-1, dtypes.float64: float('nan')})
+
+    t({dtypes.int8:  -2**7,  dtypes.uint8:  2**7})
+    t({dtypes.int16: -2**15, dtypes.uint16: 2**15})
+    t({dtypes.int32: -2**31, dtypes.uint32: 2**31})
+    t({dtypes.int64: -2**63, dtypes.uint64: 2**63})
+
+    t({dtypes.int16: 13496, dtypes.uint16: 13496, dtypes.float16: 0.294921875})
+    t({dtypes.int32: 1050081145, dtypes.uint32: 1050081145, dtypes.float32: 0.29485681653022766})
+    t({dtypes.int64: 4598983288165178391, dtypes.uint64: 4598983288165178391, dtypes.float64: 0.29485681936461233})
+
+  def test_vec_bitcast(self):
+    with Context(SPEC=0):
+      r = full_rewrite_to_sink(UOp.const(dtypes.int32.vec(3), (-1, -2**31, 75)).bitcast(dtypes.uint32.vec(3)).sink()).src[0]
+    self.assertEqual(r.op, Ops.VECTORIZE)
+    self.assertEqual(r.dtype, dtypes.uint32.vec(3))
+    self.assertEqual(tuple(x.arg for x in r.src), (2**32-1, 2**31, 75))
+
+# folds advance indexing into basic indexing
+class TestIndexingConstFolding(unittest.TestCase):
+  def test_scalar_index(self):
+    t = Tensor.arange(16).float().reshape(1,1,4,4).realize()
+    _check_ast_count(1, t[:,:,Tensor(1),:])
+    _check_ast_count(1, t[:,:,Tensor(1)+2,:])
+    _check_ast_count(1, t[:,:,Tensor(1),Tensor(0)])
+
+  def test_const_tensor_index(self):
+    # TODO: these can be 0, implement const tensor folded indexing
+    t = Tensor.arange(16).float().reshape(1,1,4,4).realize()
+    _check_ast_count(1, t[:,:,Tensor.ones(2,1,dtype=dtypes.int),:])
+    _check_ast_count(1, t[:,:,Tensor.ones(1,2,dtype=dtypes.int)+2,:])
+    _check_ast_count(1, t[:,:,Tensor.ones(1,1,dtype=dtypes.int),Tensor.zeros(2,1,2,dtype=dtypes.int)])
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_device.py

@@ -102,8 +102,8 @@ class TestCompiler(unittest.TestCase):
 
 class TestRunAsModule(unittest.TestCase):
   def test_module_runs(self):
-    out = '\n'.join(enumerate_devices_str())
-    self.assertIn("CPU", out) # for sanity check
+    cpu_line = [l for l in enumerate_devices_str() if "CPU" in l][0]
+    self.assertIn("PASS", cpu_line, f"expected CPU to PASS, got: {cpu_line}")
 
 if __name__ == "__main__":
   unittest.main()

test/null/test_dtype_spec.py

@@ -0,0 +1,413 @@
+import unittest, math, struct, operator
+from tinygrad.tensor import Tensor, dtypes
+from tinygrad.dtype import DTYPES_DICT, truncate, float_to_fp16, float_to_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
+from tinygrad.device import is_dtype_supported
+
+from tinygrad.helpers import getenv
+from hypothesis import given, settings, strategies as strat
+import numpy as np
+import torch
+
+settings.register_profile("my_profile", max_examples=50, deadline=None, derandomize=getenv("DERANDOMIZE_CI", False))
+settings.load_profile("my_profile")
+
+core_dtypes = list(DTYPES_DICT.values())
+dtype_ints = [dt for dt in core_dtypes if dtypes.is_int(dt) and is_dtype_supported(dt)]
+dtype_floats = [dt for dt in core_dtypes if dtypes.is_float(dt) and is_dtype_supported(dt)]
+
+FP8E4M3_MAX = 448.0
+FP8E5M2_MAX = 57344.0
+
+def u32_to_f32(u): return struct.unpack('f', struct.pack('I', u))[0]
+def f32_to_u32(f): return struct.unpack('I', struct.pack('f', f))[0]
+
+class TestHelpers(unittest.TestCase):
+  signed_ints = (dtypes.int8, dtypes.int16, dtypes.int32, dtypes.int64)
+  uints = (dtypes.uint8, dtypes.uint16, dtypes.uint32, dtypes.uint64)
+  floats = (dtypes.float16, dtypes.float32, dtypes.float64)
+
+  @given(strat.sampled_from(signed_ints+uints), strat.integers(min_value=1, max_value=8))
+  def test_is_int(self, dtype, amt):
+    assert dtypes.is_int(dtype.vec(amt) if amt > 1 else dtype)
+    assert not dtypes.is_float(dtype.vec(amt) if amt > 1 else dtype)
+
+  @given(strat.sampled_from(uints), strat.integers(min_value=1, max_value=8))
+  def test_is_unsigned_uints(self, dtype, amt):
+    assert dtypes.is_unsigned(dtype.vec(amt) if amt > 1 else dtype)
+
+  @given(strat.sampled_from(signed_ints), strat.integers(min_value=1, max_value=8))
+  def test_is_unsigned_signed_ints(self, dtype, amt):
+    assert not dtypes.is_unsigned(dtype.vec(amt) if amt > 1 else dtype)
+
+  @given(strat.sampled_from(floats), strat.integers(min_value=1, max_value=8))
+  def test_is_float(self, dtype, amt):
+    assert dtypes.is_float(dtype.vec(amt) if amt > 1 else dtype)
+    assert not dtypes.is_int(dtype.vec(amt) if amt > 1 else dtype)
+    assert not dtypes.is_unsigned(dtype.vec(amt) if amt > 1 else dtype)
+
+  def test_bf16_is_float(self):
+    assert dtypes.is_float(dtypes.bfloat16)
+
+  def test_fp8s_are_float(self):
+    assert dtypes.is_float(dtypes.fp8e4m3)
+    assert dtypes.is_float(dtypes.fp8e5m2)
+
+  @given(strat.sampled_from([d for d in DTYPES_DICT.values() if dtypes.is_float(d) or dtypes.is_int(d)]), strat.integers(min_value=2, max_value=8))
+  def test_scalar(self, dtype, amt):
+    assert dtype.vec(amt).scalar() == dtype
+
+  def test_from_py(self):
+    assert dtypes.from_py(True) == dtypes.bool
+    assert dtypes.from_py(2) == dtypes.default_int
+    assert dtypes.from_py(3.0) == dtypes.default_float
+    assert dtypes.from_py([]) == dtypes.default_float
+    assert dtypes.from_py(()) == dtypes.default_float
+    assert dtypes.from_py([True]) == dtypes.bool
+    assert dtypes.from_py([True, 2]) == dtypes.default_int
+    assert dtypes.from_py([True, 3.0]) == dtypes.default_float
+    assert dtypes.from_py([2, 3.0]) == dtypes.default_float
+    assert dtypes.from_py([True, 2, 3.0]) == dtypes.default_float
+    with self.assertRaises(RuntimeError): dtypes.from_py(None)
+    with self.assertRaises(RuntimeError): dtypes.from_py([None])
+    with self.assertRaises(RuntimeError): dtypes.from_py({})
+    with self.assertRaises(RuntimeError): dtypes.from_py(set())
+
+  def test_dtype_range(self):
+    for dt in core_dtypes:
+      if dtypes.is_float(dt):
+        np.testing.assert_equal(dtypes.min(dt), -math.inf)
+        np.testing.assert_equal(dtypes.max(dt), math.inf)
+        np.testing.assert_equal(dt.min, -math.inf)
+        np.testing.assert_equal(dt.max, math.inf)
+      elif dtypes.is_int(dt):
+        info = np.iinfo(_to_np_dtype(dt))
+        np.testing.assert_equal(dtypes.min(dt), info.min)
+        np.testing.assert_equal(dtypes.max(dt), info.max)
+        np.testing.assert_equal(dt.min, info.min)
+        np.testing.assert_equal(dt.max, info.max)
+      else:
+        assert dt == dtypes.bool, dt
+        np.testing.assert_equal(dtypes.min(dt), False)
+        np.testing.assert_equal(dtypes.max(dt), True)
+        np.testing.assert_equal(dt.min, False)
+        np.testing.assert_equal(dt.max, True)
+
+  def test_dtype_range_vec(self):
+    for dt in core_dtypes:
+      self.assertEqual(dt.min, dt.vec(4).min)
+      self.assertEqual(dt.max, dt.vec(4).max)
+
+  def test_float_to_fp16(self):
+    self.assertEqual(float_to_fp16(1), 1)
+    self.assertEqual(float_to_fp16(65504), 65504)
+    self.assertEqual(float_to_fp16(65519.999), 65504)
+    self.assertEqual(float_to_fp16(65520), math.inf)
+    self.assertEqual(float_to_fp16(1e-8), 0.0)
+    self.assertEqual(float_to_fp16(-65504), -65504)
+    self.assertEqual(float_to_fp16(-65519.999), -65504)
+    self.assertEqual(float_to_fp16(-65520), -math.inf)
+    self.assertTrue(math.isnan(float_to_fp16(math.nan)))
+
+  def test_float_to_bf16(self):
+    max_bf16 = torch.finfo(torch.bfloat16).max
+    for a in [1, 1.1, 1234, 23456, -777.777, max_bf16, max_bf16 * 1.00001, -max_bf16, -max_bf16 * 1.00001, math.inf, -math.inf]:
+      self.assertEqual(float_to_bf16(a), torch.tensor([a], dtype=torch.bfloat16).item())
+    self.assertTrue(math.isnan(float_to_bf16(math.nan)))
+
+  def test_float_to_bf16_nan(self):
+    patterns = [0x7FC00001, 0xFFC00001, 0x7F800001, 0xFF800001, 0x7FFFFFFF, 0xFFFFFFFF]
+    for u in patterns:
+      x = u32_to_f32(u)
+      y = float_to_bf16(x)
+      t = torch.tensor([x], dtype=torch.bfloat16).item()
+      self.assertTrue(math.isnan(y))
+      self.assertTrue(math.isnan(t))
+
+  def test_float_to_bf16_round(self):
+    uppers = [0x3f800000, 0x41230000, 0xC1460000]
+    for upper in uppers:
+      base = upper & 0xFFFF0000
+      base_f32 = u32_to_f32(base)
+      base_f32_round_up = u32_to_f32(base + 0x00010000)
+
+      x = u32_to_f32(base | 0x00007000)
+      self.assertEqual(float_to_bf16(x), base_f32)
+      self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32)
+
+      x = u32_to_f32(base | 0x0000C000)
+      self.assertEqual(float_to_bf16(x), base_f32_round_up)
+      self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32_round_up)
+
+      if ((upper >> 16) & 1) == 0:
+        x = u32_to_f32(base | 0x00008000)
+        self.assertEqual(float_to_bf16(x), base_f32)
+        self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32)
+      else:
+        x = u32_to_f32(base | 0x00008000)
+        self.assertEqual(float_to_bf16(x), base_f32_round_up)
+        self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32_round_up)
+
+  def test_float_to_bf16_boundary(self):
+    base = 0x7F7F0000
+    inf_u32 = 0x7F800000
+
+    x = u32_to_f32(base | 0x00007FFF)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), base)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), base)
+
+    x = u32_to_f32(base | 0x0000C000)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), inf_u32)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), inf_u32)
+
+    x = u32_to_f32(base | 0x00008000)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), inf_u32)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), inf_u32)
+
+  @given(strat.floats(width=32, allow_subnormal=True, allow_nan=True, allow_infinity=True))
+  def test_truncate_fp8e4m3(self, x):
+    if math.isnan(x): np.testing.assert_equal(truncate[dtypes.fp8e4m3](x), x)
+    elif math.isinf(x): np.testing.assert_equal(truncate[dtypes.fp8e4m3](x), math.copysign(math.nan, x))
+    elif x > FP8E4M3_MAX: np.testing.assert_equal(truncate[dtypes.fp8e4m3](x), FP8E4M3_MAX)
+    elif x < -FP8E4M3_MAX: np.testing.assert_equal(truncate[dtypes.fp8e4m3](x), -FP8E4M3_MAX)
+    else: np.testing.assert_equal(truncate[dtypes.fp8e4m3](x), torch.tensor(x, dtype=torch.float8_e4m3fn).float().item())
+
+  @given(strat.floats(width=32, allow_subnormal=True, allow_nan=True, allow_infinity=True))
+  def test_truncate_fp8e5m2(self, x):
+    if math.isnan(x): np.testing.assert_equal(truncate[dtypes.fp8e5m2](x), x)
+    elif math.isinf(x): np.testing.assert_equal(truncate[dtypes.fp8e5m2](x), x)
+    elif x > FP8E5M2_MAX: np.testing.assert_equal(truncate[dtypes.fp8e5m2](x), FP8E5M2_MAX)
+    elif x < -FP8E5M2_MAX: np.testing.assert_equal(truncate[dtypes.fp8e5m2](x), -FP8E5M2_MAX)
+    else: np.testing.assert_equal(truncate[dtypes.fp8e5m2](x), torch.tensor(x, dtype=torch.float8_e5m2).float().item())
+
+class TestTypePromotion(unittest.TestCase):
+  @given(strat.sampled_from(core_dtypes))
+  def test_self_promo_to_self(self, dtype):
+    assert least_upper_dtype(dtype) == dtype
+    assert least_upper_dtype(dtype, dtype) == dtype
+    assert least_upper_dtype(dtype, dtype, dtype) == dtype
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes))
+  def test_promo_resulted_higher_than_inputs(self, dtype1, dtype2):
+    result = least_upper_dtype(dtype1, dtype2)
+    assert not (result < dtype1) and not (result < dtype2)
+
+  def test_dtype_promo(self):
+    assert least_upper_dtype(dtypes.bool, dtypes.int8) == dtypes.int8
+    assert least_upper_dtype(dtypes.int8, dtypes.uint8) == dtypes.int16
+    assert least_upper_dtype(dtypes.uint8, dtypes.int16) == dtypes.int16
+    assert least_upper_dtype(dtypes.int16, dtypes.uint16) == dtypes.int32
+    assert least_upper_dtype(dtypes.uint16, dtypes.int32) == dtypes.int32
+    assert least_upper_dtype(dtypes.int32, dtypes.uint32) == dtypes.int64
+    assert least_upper_dtype(dtypes.uint32, dtypes.int64) == dtypes.int64
+    assert least_upper_dtype(dtypes.int64, dtypes.uint64) == dtypes.uint64
+    assert least_upper_dtype(dtypes.float16, dtypes.float32) == dtypes.float32
+    assert least_upper_dtype(dtypes.float32, dtypes.float64) == dtypes.float64
+
+    assert least_upper_dtype(dtypes.bool, dtypes.float32) == dtypes.float32
+    assert least_upper_dtype(dtypes.bool, dtypes.float64) == dtypes.float64
+    assert least_upper_dtype(dtypes.float16, dtypes.int64) == dtypes.float16
+    assert least_upper_dtype(dtypes.float16, dtypes.uint64) == dtypes.float16
+    assert least_upper_dtype(dtypes.fp8e4m3, dtypes.fp8e5m2) == dtypes.half
+    assert least_upper_dtype(dtypes.fp8e4m3, dtypes.bfloat16) == dtypes.bfloat16
+    assert least_upper_dtype(dtypes.fp8e5m2, dtypes.bfloat16) == dtypes.bfloat16
+    assert least_upper_dtype(dtypes.fp8e4m3, dtypes.float16) == dtypes.float16
+    assert least_upper_dtype(dtypes.fp8e5m2, dtypes.float16) == dtypes.float16
+    assert least_upper_dtype(dtypes.fp8e4m3, dtypes.int64) == dtypes.fp8e4m3
+    assert least_upper_dtype(dtypes.fp8e4m3, dtypes.uint64) == dtypes.fp8e4m3
+    assert least_upper_dtype(dtypes.fp8e5m2, dtypes.int64) == dtypes.fp8e5m2
+    assert least_upper_dtype(dtypes.fp8e5m2, dtypes.uint64) == dtypes.fp8e5m2
+
+class TestTypeSpec(unittest.TestCase):
+  def setUp(self):
+    self.old_default_int, self.old_default_float = dtypes.default_int, dtypes.default_float
+  def tearDown(self):
+    dtypes.default_int, dtypes.default_float = self.old_default_int, self.old_default_float
+
+  def test_set_dtype_default(self):
+    for default_int in [dtypes.int8, dtypes.int16, dtypes.int32, dtypes.int64]:
+      dtypes.default_int = default_int
+      assert dtypes.default_int == default_int
+
+    for default_float in [*dtypes.fp8s, dtypes.float16, dtypes.bfloat16, dtypes.float32, dtypes.float64]:
+      dtypes.default_float = default_float
+      assert dtypes.default_float == default_float
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from([operator.gt, operator.ge, operator.le, operator.lt, operator.eq, operator.ne]))
+  def test_bool_ops(self, dtype, op):
+    assert op(Tensor.ones(4, 4, dtype=dtype), Tensor.ones(4, 4, dtype=dtype)).dtype == dtypes.bool
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from(dtype_ints), strat.sampled_from(dtype_floats))
+  def test_functions_return_index(self, dtype, default_int, default_float):
+    dtypes.default_int, dtypes.default_float = default_int, default_float
+    assert Tensor([0, 1], dtype=dtype).argmax().dtype == dtypes.int32
+    assert Tensor([0, 1], dtype=dtype).argmin().dtype == dtypes.int32
+    assert Tensor([0, 1], dtype=dtype).multinomial().dtype == dtypes.int32
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from(dtype_ints))
+  def test_tensor_indexing_returns_same_dtype(self, data_dtype, indices_dtype):
+    X_data =  Tensor.ones(60000, 1, 28, 28, dtype=data_dtype)
+    indices =  Tensor.randint(512, high=X_data.shape[0]).cast(indices_dtype)
+    assert X_data[indices].dtype == X_data.dtype
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from(dtype_ints))
+  def test_gather_returns_same_dtype(self, data_dtype, indices_dtype):
+    X_data = Tensor([[1, 0], [0, 1]], dtype=data_dtype)
+    indices = Tensor([[0, 0], [1, 0]], dtype=indices_dtype)
+    assert X_data.gather(0, indices).dtype == X_data.dtype
+    assert X_data.gather(1, indices).dtype == X_data.dtype
+
+  @given(strat.sampled_from(dtype_floats), strat.sampled_from(dtype_floats))
+  def test_attention_returns_same_dtype(self, data_dtype, default_float):
+    dtypes.default_float = default_float
+    query = Tensor.rand(32, 8, 128, 64, dtype=data_dtype)
+    key = Tensor.rand(32, 8, 128, 64, dtype=data_dtype)
+    value = Tensor.rand(32, 8, 128, 64, dtype=data_dtype)
+    mask = (Tensor.rand(32, 8, 128, 128) < 0.5)
+    assert query.scaled_dot_product_attention(key, value, is_causal=True).dtype == data_dtype
+    assert query.scaled_dot_product_attention(key, value, is_causal=True, dropout_p=0.3).dtype == data_dtype
+    assert query.scaled_dot_product_attention(key, value, is_causal=False).dtype == data_dtype
+    assert query.scaled_dot_product_attention(key, value, attn_mask=mask).dtype == data_dtype
+
+class TestAutoCastType(unittest.TestCase):
+  def setUp(self):
+    self.old_default_int, self.old_default_float = dtypes.default_int, dtypes.default_float
+  def tearDown(self):
+    dtypes.default_int, dtypes.default_float = self.old_default_int, self.old_default_float
+
+  @given(strat.sampled_from(dtype_floats), strat.sampled_from(dtype_floats))
+  def test_least_upper_float_input_is_float(self, input_dtype, default_float):
+    dtypes.default_float = default_float
+    self.assertEqual(least_upper_float(input_dtype), input_dtype)
+
+  @given(strat.sampled_from(dtype_ints), strat.sampled_from(dtype_floats))
+  def test_least_upper_float_input_is_int(self, input_dtype, default_float):
+    dtypes.default_float = default_float
+    self.assertEqual(least_upper_float(input_dtype), default_float)
+
+  @given(strat.sampled_from(core_dtypes))
+  def test_broadcast_scalar(self, dt):
+    assert (Tensor.ones(4, 4, dtype=dt) + 2.3).dtype == (dt if dtypes.is_float(dt) else dtypes.default_float)
+    assert (Tensor.ones(4, 4, dtype=dt) + 2).dtype == (dt if dtypes.is_float(dt) or dtypes.is_int(dt) else dtypes.default_int)
+    assert (Tensor.ones(4, 4, dtype=dt) + True).dtype == dt
+
+  @given(strat.sampled_from(dtype_floats))
+  def test_int_div_int(self, default_float):
+    dtypes.default_float = default_float
+    self.assertEqual(Tensor([1]).div(Tensor([2])).dtype, default_float)
+
+  def test_sum(self):
+    assert (Tensor([0, 1], dtype=dtypes.bool)).sum().dtype == dtypes.int32
+    assert (Tensor([0, 1], dtype=dtypes.int8)).sum().dtype == dtypes.int32
+    assert (Tensor([0, 1], dtype=dtypes.int16)).sum().dtype == dtypes.int32
+    assert (Tensor([0, 1], dtype=dtypes.int32)).sum().dtype == dtypes.int32
+    assert (Tensor([0, 1], dtype=dtypes.int64)).sum().dtype == dtypes.int64
+    assert (Tensor([0, 1], dtype=dtypes.uint8)).sum().dtype == dtypes.uint32
+    assert (Tensor([0, 1], dtype=dtypes.uint16)).sum().dtype == dtypes.uint32
+    assert (Tensor([0, 1], dtype=dtypes.uint32)).sum().dtype == dtypes.uint32
+    assert (Tensor([0, 1], dtype=dtypes.uint64)).sum().dtype == dtypes.uint64
+    assert (Tensor([0, 1], dtype=dtypes.fp8e4m3)).sum().dtype == dtypes.fp8e4m3
+    assert (Tensor([0, 1], dtype=dtypes.fp8e5m2)).sum().dtype == dtypes.fp8e5m2
+    assert (Tensor([0, 1], dtype=dtypes.float16)).sum().dtype == dtypes.float16
+    assert (Tensor([0, 1], dtype=dtypes.bfloat16)).sum().dtype == dtypes.bfloat16
+    assert (Tensor([0, 1], dtype=dtypes.float32)).sum().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.float64)).sum().dtype == dtypes.float64
+
+  def test_mean(self):
+    assert (Tensor([0, 1], dtype=dtypes.bool)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.int8)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.int16)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.int32)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.int64)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.uint8)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.uint16)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.uint32)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.uint64)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.fp8e4m3)).mean().dtype == dtypes.fp8e4m3
+    assert (Tensor([0, 1], dtype=dtypes.fp8e5m2)).mean().dtype == dtypes.fp8e5m2
+    assert (Tensor([0, 1], dtype=dtypes.float16)).mean().dtype == dtypes.float16
+    assert (Tensor([0, 1], dtype=dtypes.bfloat16)).mean().dtype == dtypes.bfloat16
+    assert (Tensor([0, 1], dtype=dtypes.float32)).mean().dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.float64)).mean().dtype == dtypes.float64
+
+  def test_cumsum(self):
+    assert (Tensor([0, 1], dtype=dtypes.bool)).cumsum(0).dtype == dtypes.int32
+    assert (Tensor([0, 1], dtype=dtypes.int8)).cumsum(0).dtype == dtypes.int32
+    assert (Tensor([0, 1], dtype=dtypes.int16)).cumsum(0).dtype == dtypes.int32
+    assert (Tensor([0, 1], dtype=dtypes.int32)).cumsum(0).dtype == dtypes.int32
+    assert (Tensor([0, 1], dtype=dtypes.int64)).cumsum(0).dtype == dtypes.int64
+    assert (Tensor([0, 1], dtype=dtypes.uint8)).cumsum(0).dtype == dtypes.uint32
+    assert (Tensor([0, 1], dtype=dtypes.uint16)).cumsum(0).dtype == dtypes.uint32
+    assert (Tensor([0, 1], dtype=dtypes.uint32)).cumsum(0).dtype == dtypes.uint32
+    assert (Tensor([0, 1], dtype=dtypes.uint64)).cumsum(0).dtype == dtypes.uint64
+    assert (Tensor([0, 1], dtype=dtypes.fp8e4m3)).cumsum(0).dtype == dtypes.fp8e4m3
+    assert (Tensor([0, 1], dtype=dtypes.fp8e5m2)).cumsum(0).dtype == dtypes.fp8e5m2
+    assert (Tensor([0, 1], dtype=dtypes.float16)).cumsum(0).dtype == dtypes.float16
+    assert (Tensor([0, 1], dtype=dtypes.bfloat16)).cumsum(0).dtype == dtypes.bfloat16
+    assert (Tensor([0, 1], dtype=dtypes.float32)).cumsum(0).dtype == dtypes.float32
+    assert (Tensor([0, 1], dtype=dtypes.float64)).cumsum(0).dtype == dtypes.float64
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes))
+  def test_matmul(self, dt1, dt2, acc_dt):
+    t1 = Tensor([0, 1], dtype=dt1)
+    t2 = Tensor([0, 1], dtype=dt2)
+    self.assertEqual(t1.matmul(t2).dtype, least_upper_dtype(t1.dtype, t2.dtype))
+    self.assertEqual(t1.matmul(t2, dtype=acc_dt).dtype, acc_dt)
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes))
+  def test_linear(self, dt1, dt2, dt3, acc_dt):
+    x = Tensor([0, 1], dtype=dt1)
+    w = Tensor([0, 1], dtype=dt2)
+    b = Tensor([0, 1], dtype=dt3)
+    self.assertEqual(x.linear(w).dtype, least_upper_dtype(x.dtype, w.dtype))
+    self.assertEqual(x.linear(w, b).dtype, least_upper_dtype(least_upper_dtype(x.dtype, w.dtype), b.dtype))
+    self.assertEqual(x.linear(w, dtype=acc_dt).dtype, acc_dt)
+    self.assertEqual(x.linear(w, b, dtype=acc_dt).dtype, acc_dt)
+
+  @staticmethod
+  def check_where_alternate_input_other(input_, other, data_type):
+    assert (Tensor([True, False]).where(input_, other)).dtype == data_type
+    assert (Tensor([True, False]).where(other, input_)).dtype == data_type
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes))
+  def test_where_no_scalar(self, dt1, dt2):
+    self.check_where_alternate_input_other(Tensor(2, dtype=dt1), Tensor(3, dtype=dt2), least_upper_dtype(dt1, dt2))
+
+  @given(strat.sampled_from(core_dtypes))
+  def test_where_one_scalar(self, dt):
+    t = Tensor(2, dtype=dt)
+    self.check_where_alternate_input_other(t, 3.2, (dt if dtypes.is_float(dt) else dtypes.default_float))
+    self.check_where_alternate_input_other(t, 3, (dt if dtypes.is_float(dt) or dtypes.is_int(dt) else dtypes.default_int))
+    self.check_where_alternate_input_other(t, True, dt)
+
+  def test_where_two_scalars(self):
+    self.check_where_alternate_input_other(3.1, 3.2, dtypes.default_float)
+    self.check_where_alternate_input_other(3.1, 3, dtypes.default_float)
+    self.check_where_alternate_input_other(3.1, True, dtypes.default_float)
+    self.check_where_alternate_input_other(3, 2, dtypes.default_int)
+    self.check_where_alternate_input_other(3, True, dtypes.default_int)
+    self.check_where_alternate_input_other(False, True, dtypes.bool)
+
+  @given(strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes))
+  def test_maximum(self, dt1, dt2):
+    assert Tensor([0, 1, 2], dtype=dt1).maximum(Tensor([2, 0, 5], dtype=dt2)).dtype == least_upper_dtype(dt1, dt2)
+
+  @given(strat.sampled_from(core_dtypes))
+  def test_maximum_const(self, dt):
+    assert Tensor([1, 2], dtype=dt).maximum(3.1).dtype == (dt if dtypes.is_float(dt) else dtypes.default_float)
+    assert Tensor([1, 2], dtype=dt).maximum(3).dtype == (dt if dtypes.is_float(dt) or dtypes.is_int(dt) else dtypes.default_int)
+    assert Tensor([1, 2], dtype=dt).maximum(True).dtype == dt
+
+  def test_div(self):
+    assert (Tensor([1, 2], dtype=dtypes.int32) / Tensor([2, 2], dtype=dtypes.int32)).dtype == dtypes.default_float
+    assert (Tensor([1, 2], dtype=dtypes.int16) / Tensor([2, 2], dtype=dtypes.int32)).dtype == dtypes.default_float
+    assert (Tensor([1, 2], dtype=dtypes.float32) / Tensor([2, 2], dtype=dtypes.float16)).dtype == dtypes.float32
+    assert (Tensor([1, 2], dtype=dtypes.int32) / Tensor([2, 2], dtype=dtypes.float16)).dtype == dtypes.float16
+
+  def test_div_const(self):
+    assert (Tensor([1, 2], dtype=dtypes.int32) / 2).dtype == dtypes.default_float
+    assert (Tensor([1, 2], dtype=dtypes.int32) / 2.0).dtype == dtypes.default_float
+    assert (Tensor([1, 2], dtype=dtypes.float16) / 2).dtype == dtypes.float16
+    assert (Tensor([1, 2], dtype=dtypes.float16) / 2.0).dtype == dtypes.float16
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_gc.py

@@ -72,7 +72,6 @@ class TestGC(unittest.TestCase):
     ys = y.schedule()
     del x
     run_schedule(ys)
-    np.testing.assert_equal(y.numpy(), np.full((256,), 2))
     self.assertEqual(bufs_allocated()-init, 1)
     del y
     self.assertEqual(bufs_allocated()-init, 0)

test/null/test_gradient.py

@@ -0,0 +1,83 @@
+from typing import Callable
+import unittest, math
+import torch
+from tinygrad import Tensor
+from tinygrad.dtype import dtypes
+from tinygrad.uop.ops import UOp
+from tinygrad.gradient import compute_gradient
+
+class TestGradient(unittest.TestCase):
+  def _cmp_nan_okay(self, x, y):
+    if math.isnan(x) and math.isnan(y): return
+    self.assertAlmostEqual(x, y, places=5)
+
+  def _test_one_input_function(self, f:Callable, jf:Callable|None=None):
+    if jf is None: jf = f
+    x = UOp.variable('x', -math.inf, math.inf, dtype=dtypes.float)
+    gx = compute_gradient(f(x), UOp.const(dtypes.float, 1.0), set([x]))[x]
+
+    for val in [-5., -2.0, 0.0, 2.0, 5.]:
+      tg_out = gx.substitute({x: x.const_like(val)}).ssimplify()
+      tx = torch.tensor([val], dtype=torch.float, requires_grad=True)
+      torch_out = torch.autograd.grad(jf(tx), tx)[0].item()
+      self._cmp_nan_okay(tg_out, torch_out)
+
+  def _test_two_input_function(self, f:Callable, jf:Callable|None=None):
+    if jf is None: jf = f
+    x = UOp.variable('x', -math.inf, math.inf, dtype=dtypes.float)
+    y = UOp.variable('y', -math.inf, math.inf, dtype=dtypes.float)
+    grads = compute_gradient(f(x, y), UOp.const(dtypes.float, 1.0), set([x, y]))
+    gx, gy = grads[x], grads[y]
+
+    for valx in [-5., -2.0, 0.0, 2.0, 5.]:
+      for valy in [-5., -2.0, 0.0, 2.0, 5.]:
+        # Substitute the values into the gradient expressions
+        substitutions = {x: x.const_like(valx), y: y.const_like(valy)}
+        tg_out_x = gx.substitute(substitutions).ssimplify()
+        tg_out_y = gy.substitute(substitutions).ssimplify()
+
+        tx = torch.tensor([valx], dtype=torch.float, requires_grad=True)
+        ty = torch.tensor([valy], dtype=torch.float, requires_grad=True)
+        torch_grad = torch.autograd.grad(jf(tx, ty), [tx, ty])
+        torch_out_x, torch_out_y = [x.item() for x in torch_grad]
+
+        self._cmp_nan_okay(tg_out_x, torch_out_x)
+        self._cmp_nan_okay(tg_out_y, torch_out_y)
+
+  # unary ops unit
+  def test_recip(self): self._test_one_input_function(lambda x: 1.0/x)
+  def test_sin(self): self._test_one_input_function(lambda x: x.sin())
+  def test_sqrt(self): self._test_one_input_function(lambda x: x.sqrt())
+  def test_log2(self): self._test_one_input_function(lambda x: x.log2())
+  def test_exp2(self): self._test_one_input_function(lambda x: x.exp2())
+
+  # binary ops unit
+  def test_add(self): self._test_two_input_function(lambda x,y: x+y)
+  def test_mul(self): self._test_two_input_function(lambda x,y: x*y)
+
+  # chain rule
+  def test_chain(self): self._test_one_input_function(lambda x: x.sin().sqrt())
+  def test_chain_binop(self): self._test_two_input_function(lambda x,y: (x*y)+x*y)
+  def test_big_add_sin(self): self._test_two_input_function(lambda x,y: x.sin()+3.0/y)
+  def test_big_chain(self): self._test_two_input_function(lambda x,y: (1.0/x*y)+x*y)
+  def test_where(self): self._test_two_input_function(lambda x,y: (x<y).where(x,y), lambda x,y: torch.where(x<y,x,y))
+
+class TestRealizeMeansRealize(unittest.TestCase):
+  def test_randn_realizes(self):
+    x = Tensor.randn(2, 3, 64, 64, requires_grad=True).realize()
+    assert x.uop is not x.uop.base
+    assert x.uop.is_realized
+
+  def test_uniform_realizes(self):
+    x = Tensor.uniform(16, 3, 3, 3, requires_grad=True).realize()
+    print(x.uop)
+    assert x.uop is not x.uop.base
+    assert x.uop.is_realized
+
+  def test_uniform_gradient(self):
+    x = Tensor.uniform(16, 3, 3, 3, requires_grad=True).realize()
+    y = x * 2
+    y.sum().gradient(x)[0].realize()
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_helpers.py

@@ -0,0 +1,469 @@
+import ctypes, gzip, unittest, timeit, pickle
+from tinygrad import Variable
+from tinygrad.helpers import Context, ContextVar, argfix, colored, word_wrap, is_numpy_ndarray, mv_address, get_contraction, count, all_same
+from tinygrad.helpers import merge_dicts, strip_parens, prod, round_up, fetch, fully_flatten, from_mv, to_mv, polyN, time_to_str, cdiv, cmod, getbits
+from tinygrad.helpers import ceildiv
+from tinygrad.tensor import Tensor, get_shape
+import numpy as np
+
+VARIABLE = ContextVar("VARIABLE", 0)
+
+class TestContextVars(unittest.TestCase):
+  # Ensuring that the test does not modify variables outside the tests.
+  ctx = Context()
+  def setUp(self): TestContextVars.ctx.__enter__()
+  def tearDown(self): TestContextVars.ctx.__exit__()
+
+  def test_initial_value_is_set(self):
+    _TMP = ContextVar("_TMP", 5)
+    self.assertEqual(_TMP.value, 5)
+
+  def test_cannot_recreate(self):
+    _TMP2 = ContextVar("_TMP2", 1)
+    with self.assertRaises(RuntimeError):
+      _TMP2 = ContextVar("_TMP2", 2)
+
+  def test_new_var_inside_context(self):
+    with Context(VARIABLE=1):
+      _TMP3 = ContextVar("_TMP3", 1)
+    with self.assertRaises(RuntimeError):
+      _TMP3 = ContextVar("_TMP3", 2)
+
+  def test_value_across_modules(self):
+    # Mocking module import by invoking the code but not in our globals().
+    exec('from tinygrad.helpers import ContextVar;C = ContextVar("C", 13)', {}) # pylint:disable=exec-used
+    # It should not matter that the first creation was in another module.
+    with self.assertRaises(RuntimeError):
+      _C = ContextVar("C", 0)
+
+  def test_assignment_across_modules(self):
+    B = ContextVar("B", 1)
+    # local assignment
+    B.value = 2
+    self.assertEqual(B.value, 2)
+    with self.assertRaises(RuntimeError):
+      # Assignment in another module.
+      exec('from tinygrad.helpers import ContextVar;B = ContextVar("B", 0);B.value = 3;', {}) # pylint:disable=exec-used
+
+  def test_context_assignment(self):
+    with Context(VARIABLE=1):
+      self.assertEqual(VARIABLE.value, 1)
+    self.assertEqual(VARIABLE.value, 0)
+
+  def test_unknown_param_to_context(self):
+    with self.assertRaises(KeyError):
+      with Context(SOMETHING_ELSE=1):
+        pass
+
+  def test_nested_context(self):
+    with Context(VARIABLE=1):
+      with Context(VARIABLE=2):
+        MORE = ContextVar("MORE", 2)
+        with Context(VARIABLE=3, MORE=3):
+          self.assertEqual(VARIABLE.value, 3)
+          self.assertEqual(MORE.value, 3)
+        self.assertEqual(VARIABLE.value, 2)
+        self.assertEqual(MORE.value, 2)
+      self.assertEqual(VARIABLE.value, 1)
+      self.assertEqual(MORE.value, 2)  # TODO: should this raise?
+    self.assertEqual(VARIABLE.value, 0)
+
+  def test_decorator(self):
+    @Context(VARIABLE=1, DEBUG=4)
+    def test():
+      self.assertEqual(VARIABLE.value, 1)
+
+    self.assertEqual(VARIABLE.value, 0)
+    test()
+    self.assertEqual(VARIABLE.value, 0)
+
+  def test_context_exit_reverts_updated_values(self):
+    D = ContextVar("D", 1)
+    D.value = 2
+    with Context(D=3):
+      ...
+    assert D.value == 2, f"Expected D to be 2, but was {D.value}. Indicates that Context.__exit__ did not restore to the correct value."
+
+class TestAllSame(unittest.TestCase):
+  def test_empty(self): self.assertTrue(all_same([]))
+  def test_single(self): self.assertTrue(all_same([1]))
+  def test_same(self): self.assertTrue(all_same([1, 1, 1]))
+  def test_different(self): self.assertFalse(all_same([1, 2, 1]))
+
+class TestMergeDicts(unittest.TestCase):
+  def test_merge_dicts(self):
+    a = {"a": 1, "b": 2}
+    b = {"a": 1, "c": 3}
+    c = {}
+    d = {"a": 2, "b": 2}
+    assert merge_dicts([a, b]) == {"a": 1, "b": 2, "c": 3}
+    assert merge_dicts([a, c]) == a
+    assert merge_dicts([a, b, c]) == {"a": 1, "b": 2, "c": 3}
+    with self.assertRaises(RuntimeError):
+      merge_dicts([a, d])
+
+class TestStripParens(unittest.TestCase):
+  def test_simple(self): self.assertEqual("1+2", strip_parens("(1+2)"))
+  def test_nested(self): self.assertEqual("1+(2+3)", strip_parens("(1+(2+3))"))
+  def test_casted_no_strip(self): self.assertEqual("(int)(1+2)", strip_parens("(int)(1+2)"))
+  def test_unmatched_parens(self): self.assertEqual("((c35+c39>>23&255)+-127).cast(dtypes.float)",
+    strip_parens("((c35+c39>>23&255)+-127).cast(dtypes.float)"))
+  def test_single_paren_left(self): self.assertEqual("(abc", strip_parens("(abc"))
+  def test_single_paren_right(self): self.assertEqual("abc)", strip_parens("abc)"))
+  def test_parens_at_different_depths(self): self.assertEqual("(a+(b))*(c)", strip_parens("(a+(b))*(c)"))
+
+class TestProd(unittest.TestCase):
+  def test_empty(self): self.assertEqual(1, prod(tuple()))
+  def test_ints(self): self.assertEqual(30, prod((2, 3, 5)))
+  def test_variable(self): self.assertEqual("(a*12)", prod((Variable("a", 1, 5), 3, 4)).render())
+  def test_variable_order(self): self.assertEqual("(a*12)", prod((3, 4, Variable("a", 1, 5))).render())
+
+class TestRoundUp(unittest.TestCase):
+  def test_round_up(self):
+    self.assertEqual(round_up(-3,4), 0)
+    self.assertEqual(round_up(-4,4), -4)
+    self.assertEqual(round_up(6,4), 8)
+    self.assertEqual(round_up(8,4), 8)
+    self.assertEqual(round_up(232, 24984), 24984)
+    self.assertEqual(round_up(24984, 232), 25056)
+
+class TestCeilDiv(unittest.TestCase):
+  def test_int(self):
+    self.assertEqual(ceildiv(10, 3), 4)
+    self.assertEqual(ceildiv(9, 3), 3)
+    self.assertEqual(ceildiv(0, 5), 0)
+    self.assertEqual(ceildiv(1, 5), 1)
+  def test_symbolic(self):
+    # tests that ceildiv with UOp uses (num + amt - 1) // amt formula for non-negative num
+    v = Variable('v', 0, 100)
+    result = ceildiv(v, 6)
+    self.assertEqual(result.render(), "((v+5)//6)")
+  def test_symbolic_negative_offset(self):
+    # tests ceildiv(v-5, 6) which is used in conv2d output shape
+    # old implementation incorrectly simplified -(x//-y) to ((v+1)//6-1) for v-5
+    # new implementation uses (v-5+5)//6 = v//6 which is correct
+    v = Variable('v', 11, 100)
+    result = ceildiv(v - 5, 6)
+    self.assertEqual(result.render(), "(v//6)")
+
+class TestCount(unittest.TestCase):
+  def test_count_basic(self):
+    c = count(3)
+    self.assertEqual(next(c), 3)
+    self.assertEqual(next(c), 4)
+
+  def test_count_step_pickle(self):
+    c = count(1, 2)
+    self.assertEqual(next(c), 1)
+    c2 = pickle.loads(pickle.dumps(c))
+    self.assertEqual(next(c2), 3)
+
+@unittest.skip("no fetch tests because they need internet")
+class TestFetch(unittest.TestCase):
+  def test_fetch_bad_http(self):
+    self.assertRaises(Exception, fetch, 'http://www.google.com/404', allow_caching=False)
+
+  def test_fetch_small(self):
+    assert (len(fetch('https://google.com', allow_caching=False).read_bytes())>0)
+
+  def test_fetch_img(self):
+    from PIL import Image
+    img = fetch("https://avatars.githubusercontent.com/u/132956020", allow_caching=False)
+    with Image.open(img) as pimg:
+      assert pimg.size == (77, 77), pimg.size
+
+  def test_fetch_subdir(self):
+    from PIL import Image
+    img = fetch("https://avatars.githubusercontent.com/u/132956020", allow_caching=False, subdir="images")
+    with Image.open(img) as pimg:
+      assert pimg.size == (77, 77), pimg.size
+    assert img.parent.name == "images"
+
+  def test_fetch_gunzip_valid(self):
+    # compare fetch(gunzip=True) to fetch(gunzip=False) plus decompressing afterwards
+    gzip_url: str = 'https://ftp.gnu.org/gnu/gzip/gzip-1.13.tar.gz'
+    fp_gz = fetch(gzip_url, gunzip=True)
+    fp_no_gz = fetch(gzip_url, gunzip=False)
+    with open(fp_gz, 'rb') as f: content_gz = f.read()
+    with open(fp_no_gz, 'rb') as f: content_no_gz = gzip.decompress(f.read())
+    assert fp_gz.stat().st_size > fp_no_gz.stat().st_size
+    assert isinstance(content_gz, bytes) and isinstance(content_no_gz, bytes)
+    assert len(content_gz) == len(content_no_gz)
+    assert content_gz == content_no_gz
+
+  def test_fetch_gunzip_invalid(self):
+    # given a non-gzipped file, fetch(gunzip=True) fails
+    no_gzip_url: str = 'https://ftp.gnu.org/gnu/gzip/gzip-1.13.zip'
+    with self.assertRaises(gzip.BadGzipFile):
+      fetch(no_gzip_url, gunzip=True)
+
+  def test_fetch_user_agent(self):
+    fetch("https://csrc.nist.gov/CSRC/media/Projects/lightweight-cryptography/documents/finalist-round/updated-submissions/sparkle.zip",
+          allow_caching=False)
+
+  def test_fetch_half_and_full_file(self):
+    x = fetch("https://csrc.nist.gov/CSRC/media/Projects/lightweight-cryptography/documents/finalist-round/updated-submissions/sparkle.zip",
+          headers={"Range": "bytes=0-10"}).read_bytes()
+    assert len(x) == 11, f"{len(x) != 11}"
+    x = fetch("https://csrc.nist.gov/CSRC/media/Projects/lightweight-cryptography/documents/finalist-round/updated-submissions/sparkle.zip",
+          headers={"Range": "bytes=0-100"}).read_bytes()
+    assert len(x) == 101, f"{len(x) != 101}"
+
+class TestFullyFlatten(unittest.TestCase):
+  def test_fully_flatten(self):
+    self.assertEqual(fully_flatten([[1, 3], [1, 2]]), [1, 3, 1, 2])
+    self.assertEqual(fully_flatten(((1, 3), (1, 2))), [1, 3, 1, 2])
+    self.assertEqual(fully_flatten([[[1], [3]], [[1], [2]]]), [1, 3, 1, 2])
+    self.assertEqual(fully_flatten([[[[1], 2], 3], 4]), [1, 2, 3, 4])
+    self.assertEqual(fully_flatten([[1, 2, [3, 4]], [5, 6], 7]), [1, 2, 3, 4, 5, 6, 7])
+    self.assertEqual(fully_flatten([[1, "ab"], [True, None], [3.14, [5, "b"]]]), [1, "ab", True, None, 3.14, 5, "b"])
+
+  def test_fully_flatten_numpy(self):
+    self.assertEqual(fully_flatten([np.array([])]), [])
+    self.assertEqual(fully_flatten([np.array(3)]), [3])
+    self.assertEqual(fully_flatten([np.array([3])]), [3])
+    self.assertEqual(fully_flatten([np.array([[3]])]), [3])
+    self.assertEqual(fully_flatten([np.array([1, 3]), np.array([1, 2])]), [1, 3, 1, 2])
+    self.assertEqual(fully_flatten((np.array([1, 3]), np.array([1, 2]))), [1, 3, 1, 2])
+    self.assertEqual(fully_flatten([np.array([[1], [3]]), np.array([[1], [2]])]), [1, 3, 1, 2])
+    self.assertEqual(fully_flatten([[1, "ab"], [True, None], np.array([[3.14], [6.28]])]), [1, "ab", True, None, 3.14, 6.28])
+
+class TestMemoryview(unittest.TestCase):
+  def test_from_mv_to_mv(self):
+    base = memoryview(bytearray(b"\x11\x22\x33"*40))
+    ct = from_mv(base)
+    mv = to_mv(ctypes.addressof(ct), len(base))
+    mv[0] = 2
+    assert base[0] == 2
+
+  @unittest.skip("allocates tons of memory")
+  def test_to_mv(self):
+    sizes = [
+      (16, "16 B"),
+      (64, "64 B"),
+      (256, "256 B"),
+      (1024, "1 KB"),
+      (4 * 1024, "4 KB"),
+      (16 * 1024, "16 KB"),
+      (64 * 1024, "64 KB"),
+      (256 * 1024, "256 KB"),
+      (1 * 1024 * 1024, "1 MB"),
+      (10 * 1024 * 1024, "10 MB"),
+      (200 * 1024 * 1024, "200 MB"),
+    ]
+
+    for sz, label in sizes:
+      buf = np.random.randint(0, 256, sz, dtype=np.uint8)
+      ptr = buf.ctypes.data
+
+      iters = 100_000
+      t_us = timeit.timeit(lambda: to_mv(ptr, sz), number=iters) * 1e6 / iters
+      print(f"Size {label:>9} | Time: {t_us:8.3f} µs")
+
+  def test_speed_from_mv_vs_mv_address(self):
+    x = memoryview(bytearray(1))
+
+    iters = 100000
+    fmv_us = timeit.timeit(lambda: from_mv(x), number=iters) * 1e6 / iters
+    mva_us = timeit.timeit(lambda: mv_address(x), number=iters) * 1e6 / iters
+    print(f"from_mv vs mv_address: {fmv_us:8.3f} µs vs {mva_us:8.3f} µs")
+
+class TestGetContraction(unittest.TestCase):
+  def test_contraction(self):
+    r = get_contraction((1,2,3,4), (2,3,4))
+    self.assertEqual(r, [[0, 1], [2], [3]])
+
+    r = get_contraction((2,1,3,4), (2,3,4))
+    self.assertEqual(r, [[0], [1, 2], [3]])
+
+    r = get_contraction((1,2,3,1,4), (1,2,3,4))
+    self.assertEqual(r, [[], [0, 1], [2], [3, 4]])
+
+    r = get_contraction((1,2,3,1,4,1,1), (2,3,4))
+    self.assertEqual(r, [[0, 1], [2], [3, 4, 5, 6]])
+
+    r = get_contraction((1,2,3,4), (1,2,3*4))
+    self.assertEqual(r, [[], [0, 1], [2, 3]])
+
+    r = get_contraction((1,2,3,4), (2,1,3,4))
+    self.assertEqual(r, [[0, 1], [], [2], [3]])
+
+    r = get_contraction((1,2,3,4), (1,1,2*3*4,1))
+    self.assertEqual(r, [[], [], [0,1,2,3], []])
+
+    r = get_contraction((2,1,3,4), (1,2,3,4))
+    self.assertEqual(r, [[], [0], [1, 2], [3]])
+
+    r = get_contraction((1,2,3,4), (2*3*4,1,1,1))
+    self.assertEqual(r, [[0, 1, 2, 3], [], [], []])
+
+    r = get_contraction((4,4,4,4), (16,1,16))
+    self.assertEqual(r, [[0, 1], [], [2, 3]])
+
+    r = get_contraction((1,2,3,4,1,1,1), (2,3,4))
+    self.assertEqual(r, [[0, 1], [2], [3, 4, 5, 6]])
+
+    r = get_contraction((1,2,3,4), (1,2,3,4,1))
+    self.assertEqual(r, [[], [0, 1], [2], [3], []])
+
+    r = get_contraction((14,1,384,14,1,1,1,1), (1,14,384,14))
+    self.assertEqual(r, [[], [0], [1,2], [3,4,5,6,7]])
+
+    r = get_contraction((14,1,384,1,14,1,1,1,1), (1,14,384,14))
+    self.assertEqual(r, [[], [0], [1,2], [3,4,5,6,7,8]])
+
+    r = get_contraction((512, 512), (1, 1, 512, 1, 1, 1, 1, 512))
+    self.assertEqual(r, [[], [], [0], [], [], [], [], [1]])
+
+    r = get_contraction((1,2,3,4), (1,2,6,2))
+    self.assertEqual(r, None)
+
+  def test_contraction_ones(self):
+    r = get_contraction((1,), (1,1,1))
+    self.assertEqual(r, [[], [], [0]])
+
+    r = get_contraction((1,1), (1,1,1))
+    self.assertEqual(r, [[], [], [0, 1]])
+
+    r = get_contraction((1,1,1,1), (1,))
+    self.assertEqual(r, [[0,1,2,3]])
+
+    r = get_contraction((1,1,1,1), (1,1))
+    self.assertEqual(r, [[], [0,1,2,3]])
+
+    r = get_contraction((1,1,1,1), (1,1,1))
+    self.assertEqual(r, [[], [], [0,1,2,3]])
+
+    r = get_contraction((1,1,1,1), (1,1,1,1))
+    self.assertEqual(r, [[], [], [], [0,1,2,3]])
+
+class TestGetShape(unittest.TestCase):
+  def test_get_shape(self):
+    assert get_shape(2) == ()
+    assert get_shape([]) == (0,)
+    assert get_shape([[]]) == (1, 0)
+    assert get_shape([[1, 2]]) == (1, 2)
+    assert get_shape([[1, 2], (3, 4)]) == (2, 2)
+
+  def test_inhomogeneous_shape(self):
+    with self.assertRaises(ValueError): get_shape([[], [1]])
+    with self.assertRaises(ValueError): get_shape([[1, [2]], [1]])
+
+class TestPolyN(unittest.TestCase):
+  def test_float(self):
+    np.testing.assert_allclose(polyN(1.0, [1.0, -2.0, 1.0]), 0.0)
+    np.testing.assert_allclose(polyN(2.0, [1.0, -2.0, 1.0]), 1.0)
+    np.testing.assert_allclose(polyN(3.0, [1.0, -2.0, 1.0]), 4.0)
+    np.testing.assert_allclose(polyN(4.0, [1.0, -2.0, 1.0]), 9.0)
+
+  def test_uop(self):
+    from tinygrad.dtype import dtypes
+    from tinygrad.uop.ops import UOp
+    from test.helpers import eval_uop
+    np.testing.assert_allclose(eval_uop(polyN(UOp.const(dtypes.float, 1.0), [1.0, -2.0, 1.0])), 0.0)
+    np.testing.assert_allclose(eval_uop(polyN(UOp.const(dtypes.float, 2.0), [1.0, -2.0, 1.0])), 1.0)
+    np.testing.assert_allclose(eval_uop(polyN(UOp.const(dtypes.float, 3.0), [1.0, -2.0, 1.0])), 4.0)
+    np.testing.assert_allclose(eval_uop(polyN(UOp.const(dtypes.float, 4.0), [1.0, -2.0, 1.0])), 9.0)
+
+class TestTimeToStr(unittest.TestCase):
+  def test_seconds(self):           self.assertEqual("   10.01s ", time_to_str(10.01))
+  def test_boundary_sec_ms(self):   self.assertEqual("10000.00ms", time_to_str(10))
+  def test_milliseconds(self):      self.assertEqual("  500.00ms", time_to_str(0.5))
+  def test_boundary_ms_us(self):    self.assertEqual("10000.00us", time_to_str(0.01))
+  def test_microseconds(self):      self.assertEqual("  100.00us", time_to_str(0.0001))
+  def test_zero(self):              self.assertEqual("    0.00us", time_to_str(0))
+  def test_width_formatting(self):  self.assertEqual(" 10.01s ", time_to_str(10.01, w=6))
+
+class TestCStyleDivMod(unittest.TestCase):
+  def test_div_pos(self):
+    self.assertEqual(cdiv(-9, 5), -1)
+    self.assertEqual(cdiv(-4, 5), 0)
+    self.assertEqual(cdiv(0, 5), 0)
+    self.assertEqual(cdiv(4, 5), 0)
+    self.assertEqual(cdiv(9, 5), 1)
+  def test_div_neg(self):
+    self.assertEqual(cdiv(-9, -5), 1)
+    self.assertEqual(cdiv(-4, -5), 0)
+    self.assertEqual(cdiv(0, -5), 0)
+    self.assertEqual(cdiv(4, -5), 0)
+    self.assertEqual(cdiv(9, -5), -1)
+  def test_mod_pos(self):
+    self.assertEqual(cmod(-9, 5), -4)
+    self.assertEqual(cmod(-4, 5), -4)
+    self.assertEqual(cmod(0, 5), 0)
+    self.assertEqual(cmod(4, 5), 4)
+    self.assertEqual(cmod(9, 5), 4)
+  def test_mod_neg(self):
+    self.assertEqual(cmod(-9, -5), -4)
+    self.assertEqual(cmod(-4, -5), -4)
+    self.assertEqual(cmod(0, -5), 0)
+    self.assertEqual(cmod(4, -5), 4)
+    self.assertEqual(cmod(9, -5), 4)
+
+class TestGetBits(unittest.TestCase):
+  def test_low_bits(self):
+    self.assertEqual(getbits(0b11010110, 0, 3), 0b0110)
+
+  def test_high_bits(self):
+    self.assertEqual(getbits(0b11010110, 4, 7), 0b1101)
+
+  def test_middle_bits(self):
+    self.assertEqual(getbits(0b11010110, 3, 5), 0b010)
+
+  def test_full_range(self):
+    self.assertEqual(getbits(0b11010110, 0, 7), 0b11010110)
+
+  def test_single_bit(self):
+    self.assertEqual(getbits(0b100000000, 8, 8), 1)
+
+class TestArgFix(unittest.TestCase):
+  def test_none(self):
+    self.assertEqual(argfix(None), (None, ))
+    self.assertEqual(argfix(None, None), (None, None))
+  def test_positional_arguments(self):
+    self.assertEqual(argfix(1, 2, 3), (1, 2, 3))
+  def test_tuple(self):
+    self.assertEqual(argfix((1., 2., 3.)), (1., 2., 3.))
+  def test_list(self):
+    self.assertEqual(argfix([True, False]), (True, False))
+
+class TestWordWrap(unittest.TestCase):
+  def test_wrap_simple(self):
+    wrap = 10
+    st = "x"*wrap*2
+    st2 = word_wrap(st, wrap)
+    self.assertEqual(len(st2.splitlines()), 2)
+
+  def test_wrap_colored(self):
+    wrap = 10
+    st = colored("x"*wrap*2, "red")
+    st2 = word_wrap(st, wrap=wrap)
+    self.assertEqual(len(st2.splitlines()), 2)
+
+  def test_wrap_explicit_newline(self):
+    wrap = 10
+    st = "\n".join(["x"*wrap, "x"*wrap, "x"*wrap])
+    st2 = word_wrap(st, wrap=wrap)
+    self.assertEqual(len(st2.splitlines()), len(st.splitlines()))
+
+    st = "\n".join(["x"*(wrap+1), "x"*wrap, "x"*wrap])
+    st2 = word_wrap(st, wrap=wrap)
+    self.assertEqual(len(st2.splitlines()), len(st.splitlines())+1)
+
+    st = "\n".join(["x"*(wrap+1), "x"*(wrap+1), "x"*(wrap+1)])
+    st2 = word_wrap(st, wrap=wrap)
+    self.assertEqual(len(st2.splitlines()), len(st.splitlines())+3)
+
+class TestIsNumpyNdarray(unittest.TestCase):
+  def test_ndarray(self):
+    self.assertTrue(is_numpy_ndarray(np.array([1, 2, 3])))
+  def test_ndarray_tolist(self):
+    self.assertFalse(is_numpy_ndarray(np.array([1, 2, 3]).tolist()))
+  def test_list(self):
+    self.assertFalse(is_numpy_ndarray([1, 2, 3]))
+  def test_tensor(self):
+    self.assertFalse(is_numpy_ndarray(Tensor([1, 2, 3])))
+    self.assertFalse(is_numpy_ndarray(Tensor(np.array([1, 2, 3]))))
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_indexing.py

@@ -0,0 +1,143 @@
+# test cases are modified from pytorch test_indexing.py
+
+import unittest
+import numpy as np
+
+from tinygrad import Tensor, dtypes
+
+def numpy_testing_assert_equal_helper(a, b):
+  if isinstance(a, Tensor): a = a.numpy()
+  if isinstance(b, Tensor): b = b.numpy()
+  np.testing.assert_equal(a, b)
+
+class TestIndexing(unittest.TestCase):
+  def test_single_int(self):
+    v = Tensor.randn(5, 7, 3)
+    numpy_testing_assert_equal_helper(v[4].shape, (7, 3))
+
+  def test_multiple_int(self):
+    v = Tensor.randn(5, 7, 3)
+    numpy_testing_assert_equal_helper(v[4].shape, (7, 3))
+    numpy_testing_assert_equal_helper(v[4, :, 1].shape, (7,))
+
+  def test_none(self):
+    v = Tensor.randn(5, 7, 3)
+    numpy_testing_assert_equal_helper(v[None].shape, (1, 5, 7, 3))
+    numpy_testing_assert_equal_helper(v[:, None].shape, (5, 1, 7, 3))
+    numpy_testing_assert_equal_helper(v[:, None, None].shape, (5, 1, 1, 7, 3))
+    numpy_testing_assert_equal_helper(v[..., None].shape, (5, 7, 3, 1))
+
+  def test_int_indices(self):
+    v = Tensor.randn(5, 7, 3)
+    numpy_testing_assert_equal_helper(v[[0, 4, 2]].shape, (3, 7, 3))
+    numpy_testing_assert_equal_helper(v[:, [0, 4, 2]].shape, (5, 3, 3))
+    numpy_testing_assert_equal_helper(v[:, [[0, 1], [4, 3]]].shape, (5, 2, 2, 3))
+
+  def test_index_src_datatype(self):
+    src = Tensor.ones(3, 2, 4)
+    # test index
+    res = src[[0, 2, 1], :, :]
+    numpy_testing_assert_equal_helper(res.shape, src.shape)
+
+  def test_empty_slice(self):
+    x = Tensor.randn(2, 3, 4, 5)
+    y = x[:, :, :, 1]
+    z = y[:, 1:1, :]
+    numpy_testing_assert_equal_helper((2, 0, 4), z.shape)
+
+  def test_invalid_index(self):
+    x = Tensor.arange(0, 16).reshape(4, 4)
+    self.assertRaises(TypeError, lambda: x["0":"1"])
+
+  def test_out_of_bound_index(self):
+    x = Tensor.arange(0, 100).reshape(2, 5, 10)
+    self.assertRaises(IndexError, lambda: x[0, 5])
+    self.assertRaises(IndexError, lambda: x[4, 5])
+    self.assertRaises(IndexError, lambda: x[0, 1, 15])
+    self.assertRaises(IndexError, lambda: x[:, :, 12])
+
+  def test_take_along_dim(self):
+    # NOTE: the actual test logic is inside _test_against_numpy which is never called
+    # This test effectively does nothing but defines a function
+    def _test_against_numpy(t: Tensor, indices: Tensor, dim):
+      actual = t.gather(dim, indices)
+      t_np = t.numpy()
+      indices_np = indices.numpy()
+      expected = np.take_along_axis(t_np, indices_np, axis=dim)
+      numpy_testing_assert_equal_helper(actual, expected)
+
+      # TODO argsort
+      '''
+      for shape in [(3, 2), (2, 3, 5), (2, 4, 0), (2, 3, 1, 4)]:
+        for noncontiguous in [True, False]:
+          for dtype in (dtypes.float32, dtypes.int64):
+            t = make_tensor(shape, dtype=dtype, noncontiguous=noncontiguous)
+            for dim in list(range(t.ndim)) + [None]:
+              if dim is None:
+                indices = argsort(t.reshape(-1))
+              else:
+                indices = argsort(t, dim=dim)
+
+          _test_against_numpy(t, indices, dim)
+      '''
+
+      # test broadcasting
+      t = Tensor.ones((3, 4, 1))
+      indices = Tensor.ones((1, 2, 5), dtype=dtypes.int64)
+
+      _test_against_numpy(t, indices, 1)
+
+      # test empty indices
+      t = Tensor.ones((3, 4, 5))
+      indices = Tensor.ones((3, 0, 5), dtype=dtypes.int64)
+
+      _test_against_numpy(t, indices, 1)
+
+class TestNumpy(unittest.TestCase):
+  def test_index_no_floats(self):
+    a = Tensor([[[5.]]])
+
+    self.assertRaises(IndexError, lambda: a[0.0])
+    self.assertRaises(IndexError, lambda: a[0, 0.0])
+    self.assertRaises(IndexError, lambda: a[0.0, 0])
+    self.assertRaises(IndexError, lambda: a[0.0, :])
+    self.assertRaises(IndexError, lambda: a[:, 0.0])
+    self.assertRaises(IndexError, lambda: a[:, 0.0, :])
+    self.assertRaises(IndexError, lambda: a[0.0, :, :])
+    self.assertRaises(IndexError, lambda: a[0, 0, 0.0])
+    self.assertRaises(IndexError, lambda: a[0.0, 0, 0])
+    self.assertRaises(IndexError, lambda: a[0, 0.0, 0])
+    self.assertRaises(IndexError, lambda: a[-1.4])
+    self.assertRaises(IndexError, lambda: a[0, -1.4])
+    self.assertRaises(IndexError, lambda: a[-1.4, 0])
+    self.assertRaises(IndexError, lambda: a[-1.4, :])
+    self.assertRaises(IndexError, lambda: a[:, -1.4])
+    self.assertRaises(IndexError, lambda: a[:, -1.4, :])
+    self.assertRaises(IndexError, lambda: a[-1.4, :, :])
+    self.assertRaises(IndexError, lambda: a[0, 0, -1.4])
+    self.assertRaises(IndexError, lambda: a[-1.4, 0, 0])
+    self.assertRaises(IndexError, lambda: a[0, -1.4, 0])
+    # these two trigger slice internal type verification first
+    self.assertRaises(TypeError, lambda: a[0.0:, 0.0])
+    self.assertRaises(TypeError, lambda: a[0.0:, 0.0,:])
+
+  def test_none_index(self):
+    # `None` index adds newaxis
+    a = Tensor([1, 2, 3])
+    numpy_testing_assert_equal_helper(a[None].ndim, a.ndim+1)
+
+  def test_everything_returns_views(self):
+    # Before `...` would return a itself.
+    a = Tensor([5])
+
+    self.assertIs(a, a[()])
+    self.assertIs(a, a[...])
+    self.assertIs(a, a[:])
+
+  def test_broaderrors_indexing(self):
+    a = Tensor.zeros(5, 5)
+    self.assertRaises(IndexError, a.__getitem__, ([0, 1], [0, 1, 2]))
+    self.assertRaises(IndexError, a.contiguous().__setitem__, ([0, 1], [0, 1, 2]), 0)
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_linearizer_failures.py

@@ -1,6 +1,6 @@
 # ruff: noqa: E501
 import unittest
-from tinygrad.uop.ops import UOp, Ops, AxisType
+from tinygrad.uop.ops import UOp, Ops, AxisType, KernelInfo
 from tinygrad.dtype import dtypes
 from tinygrad.engine.realize import get_program
 from tinygrad.device import Device
@@ -18,7 +18,7 @@ class TestLinearizerFailures(unittest.TestCase):
     c8 = c7.index(c3)
     c9 = ((((c6+(c8*UOp.const(dtypes.float, -1.0)))*(c6+(c8*UOp.const(dtypes.float, -1.0)))).reduce(c5, arg=Ops.ADD)*UOp.const(dtypes.float, 0.000390625))+UOp.const(dtypes.float, 1e-05)).sqrt().reciprocal()
     c10 = c0.index(c3).store(c9).end(c1, c2)
-    ast = c10.sink()
+    ast = c10.sink(arg=KernelInfo())
     get_program(ast, renderer=Device[Device.DEFAULT].renderer)
 
 if __name__ == '__main__':

test/null/test_linearizer_rewrite.py

@@ -35,7 +35,7 @@ class TestLinearizerRewrite(unittest.TestCase):
     prg = get_program(ast, Device["CPU"].renderer)
     assert prg.applied_opts == (), f"expected no opts, got {prg}"
 
-    prg = get_program(ast.replace(arg=None), Device["CPU"].renderer)
+    prg = get_program(ast.replace(arg=KernelInfo()), Device["CPU"].renderer)
     assert prg.applied_opts != (), f"expected opts to apply, got {prg.applied_opts}"
 
     prg = get_program(ast.replace(arg=KernelInfo(name="custom")), Device["CPU"].renderer)

test/null/test_llm_server.py

@@ -0,0 +1,137 @@
+import unittest, threading, time
+from unittest.mock import Mock
+
+class TestLLMServer(unittest.TestCase):
+  """Integration tests using the real OpenAI client."""
+
+  @classmethod
+  def setUpClass(cls):
+    cls.mock_tok = Mock()
+    cls.mock_tok.role = Mock(return_value=[100, 101])
+    cls.mock_tok.encode = Mock(return_value=[200, 201, 202])
+    cls.mock_tok.decode = Mock(return_value="Hello")
+    cls.mock_tok.end_turn = Mock(return_value=[998])
+
+    cls.mock_model = Mock()
+    cls.mock_model.generate = Mock(side_effect=lambda ids, **kwargs: iter([300, 301, 999]))
+
+    cls.bos_id = 1
+    cls.eos_id = 999
+
+    import tinygrad.apps.llm as llm_module
+    llm_module.model = cls.mock_model
+    llm_module.tok = cls.mock_tok
+    llm_module.bos_id = cls.bos_id
+    llm_module.eos_id = cls.eos_id
+
+    from tinygrad.apps.llm import Handler
+    from tinygrad.viz.serve import TCPServerWithReuse
+
+    cls.server = TCPServerWithReuse(('127.0.0.1', 0), Handler)
+    cls.port = cls.server.server_address[1]
+    cls.server_thread = threading.Thread(target=cls.server.serve_forever, daemon=True)
+    cls.server_thread.start()
+    time.sleep(0.1)
+
+    from openai import OpenAI
+    cls.client = OpenAI(base_url=f"http://127.0.0.1:{cls.port}/v1", api_key="test")
+
+  @classmethod
+  def tearDownClass(cls):
+    cls.server.shutdown()
+    cls.server.server_close()
+
+  def test_chat_completion_stream(self):
+    stream = self.client.chat.completions.create(
+      model="test",
+      messages=[{"role": "user", "content": "Hello"}],
+      stream=True
+    )
+
+    chunks = list(stream)
+    self.assertGreater(len(chunks), 0)
+    self.assertEqual(chunks[0].choices[0].delta.role, "assistant")
+    self.assertEqual(chunks[-1].choices[0].finish_reason, "stop")
+
+  def test_openai_response_structure(self):
+    stream = self.client.chat.completions.create(
+      model="test-model",
+      messages=[{"role": "user", "content": "Test"}],
+      stream=True
+    )
+
+    for chunk in stream:
+      self.assertTrue(chunk.id.startswith("chatcmpl-"))
+      self.assertEqual(chunk.object, "chat.completion.chunk")
+      self.assertIsNotNone(chunk.choices)
+      self.assertIsNotNone(chunk.created)
+      self.assertIsInstance(chunk.created, int)
+      self.assertEqual(chunk.model, "test-model")
+
+  def test_stream_with_usage(self):
+    stream = self.client.chat.completions.create(
+      model="test",
+      messages=[{"role": "user", "content": "Hello"}],
+      stream=True,
+      stream_options={"include_usage": True}
+    )
+
+    chunks = list(stream)
+    last_chunk = chunks[-1]
+
+    self.assertIsNotNone(last_chunk.usage)
+    self.assertIsNotNone(last_chunk.usage.prompt_tokens)
+    self.assertIsNotNone(last_chunk.usage.completion_tokens)
+    self.assertIsNotNone(last_chunk.usage.total_tokens)
+
+  def test_multi_turn_conversation(self):
+    stream = self.client.chat.completions.create(
+      model="test",
+      messages=[
+        {"role": "system", "content": "You are helpful."},
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "Hi!"},
+        {"role": "user", "content": "How are you?"}
+      ],
+      stream=True
+    )
+
+    chunks = list(stream)
+    self.assertGreater(len(chunks), 0)
+    self.assertEqual(chunks[-1].choices[0].finish_reason, "stop")
+
+  def test_content_is_streamed(self):
+    stream = self.client.chat.completions.create(
+      model="test",
+      messages=[{"role": "user", "content": "Hello"}],
+      stream=True
+    )
+
+    contents = []
+    for chunk in stream:
+      if chunk.choices and chunk.choices[0].delta.content:
+        contents.append(chunk.choices[0].delta.content)
+
+    self.assertGreater(len(contents), 0)
+
+  def test_non_streaming(self):
+    resp = self.client.chat.completions.create(
+      model="test-model",
+      messages=[{"role": "user", "content": "Hello"}],
+      stream=False
+    )
+
+    self.assertTrue(resp.id.startswith("chatcmpl-"))
+    self.assertEqual(resp.object, "chat.completion")
+    self.assertEqual(resp.model, "test-model")
+    self.assertIsNotNone(resp.created)
+    self.assertEqual(len(resp.choices), 1)
+    self.assertEqual(resp.choices[0].message.role, "assistant")
+    self.assertIsNotNone(resp.choices[0].message.content)
+    self.assertEqual(resp.choices[0].finish_reason, "stop")
+    self.assertIsNotNone(resp.usage)
+    self.assertIsNotNone(resp.usage.prompt_tokens)
+    self.assertIsNotNone(resp.usage.completion_tokens)
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_multitensor.py

@@ -0,0 +1,58 @@
+import unittest
+from tinygrad import Tensor, GlobalCounters, dtypes
+
+class TestMultiRamUsage(unittest.TestCase):
+  def setUp(self):
+    self.baseline = GlobalCounters.mem_used
+    self.N = 100
+  def assertUsed(self, amt, strict=True):
+    used = GlobalCounters.mem_used - self.baseline
+    print(f"used {used} bytes")
+    if strict: self.assertEqual(used, amt)
+    else: self.assertLessEqual(used, amt)
+
+  def test_zeros(self):
+    _ = Tensor.zeros(self.N, self.N).contiguous().realize()
+    self.assertUsed(self.N*self.N*4)
+
+  def test_zeros_del(self):
+    _ = Tensor.zeros(self.N, self.N).contiguous().realize()
+    del _
+    self.assertUsed(0)
+
+  @unittest.skip("flaky")
+  def test_zeros_copy(self):
+    devices_2 = ("NULL:1", "NULL:2")
+    _ = Tensor.zeros(self.N, self.N).contiguous().to(devices_2).realize()
+    # NOTE: the first one on the DEFAULT device should be freed
+    self.assertUsed(self.N*self.N*4*2)
+
+  @unittest.skip("flaky")
+  def test_zeros_shard(self, devices=("NULL:1", "NULL:2")):
+    _ = Tensor.zeros(self.N, self.N).contiguous().shard(devices, axis=0).realize()
+    self.assertUsed(self.N*self.N*4) # sharding should not increase total ram usage
+  def test_zeros_shard_self(self): self.test_zeros_shard(("NULL:0", "NULL:1"))
+
+  @unittest.skip("flaky")
+  def test_zeros_contiguous_shard(self):
+    devices_2 = ("NULL:1", "NULL:2")
+    _ = Tensor.zeros(self.N, self.N).contiguous().shard(devices_2, axis=0).contiguous().realize()
+    self.assertUsed(self.N*self.N*4) # sharding should not increase total ram usage
+
+  def _test_matmul_half(self, dev_count:int):
+    N = 32
+    total_mem = {}
+    devs = tuple(f"NULL:{i}" for i in range(dev_count))
+    for dtype in {dtypes.float, dtypes.half}:
+      GlobalCounters.reset()
+      a = Tensor.empty((N, N), dtype=dtype, device=devs[0]).shard(devs, axis=0)
+      b = Tensor.empty((N, N), dtype=dtype, device=devs[0]).shard(devs, axis=None)
+      (a @ b).realize()
+      total_mem[dtype] = GlobalCounters.global_mem
+    self.assertEqual(total_mem[dtypes.half], total_mem[dtypes.float] // 2)
+
+  def test_matmul_half(self): self._test_matmul_half(dev_count=2)
+  def test_matmul_half_alt(self): self._test_matmul_half(dev_count=4)
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_process_replay.py

@@ -0,0 +1,34 @@
+import unittest
+from tinygrad import Tensor, Device
+from tinygrad.engine.realize import get_program
+from tinygrad.codegen.opt import Opt, OptOps
+from test.external.process_replay.process_replay import replay_get_program
+
+N = 16
+class TestProcessReplay(unittest.TestCase):
+  @classmethod
+  def setUpClass(cls):
+    cls.ast = (Tensor.empty(N, N) @ Tensor.empty(N, N)).schedule()[-1].ast
+    cls.renderer = Device[Device.DEFAULT].renderer
+
+  def test_replay_no_opts(self):
+    # opts=None means use default heuristic path
+    p = get_program(self.ast, self.renderer)
+    good, compare, _ = replay_get_program(p, self.ast, self.renderer)
+    self.assertEqual(good, compare)
+
+  def test_replay_empty_opts(self):
+    # opts=[] means explicitly apply zero opts (unoptimized)
+    p = get_program(self.ast, self.renderer, opts=[])
+    good, compare, _ = replay_get_program(p, self.ast, self.renderer, opts=[])
+    self.assertEqual(good, compare)
+
+  def test_replay_with_opt(self):
+    # opts=[Opt(...)] means apply a specific opt
+    opts = [Opt(OptOps.UPCAST, 0, 4)]
+    p = get_program(self.ast, self.renderer, opts=opts)
+    good, compare, _ = replay_get_program(p, self.ast, self.renderer, opts=opts)
+    self.assertEqual(good, compare)
+
+if __name__ == '__main__':
+  unittest.main(verbosity=2)

test/null/test_real_world.py

@@ -12,7 +12,7 @@ from test.helpers import derandomize_model
 
 from examples.gpt2 import Transformer as GPT2Transformer
 from examples.hlb_cifar10 import SpeedyResNet, hyp
-from examples.llama import Transformer as LLaMaTransformer
+from extra.models.llama import Transformer as LLaMaTransformer
 from examples.stable_diffusion import UNetModel, unet_params
 from extra.models.unet import ResBlock
 from extra.models.bert import BertForPretraining

test/null/test_rearrange_einops.py

@@ -0,0 +1,171 @@
+# modified from
+# https://github.com/arogozhnikov/einops/blob/master/tests/test_examples.py
+# https://github.com/arogozhnikov/einops/blob/master/tests/test_ops.py
+# https://github.com/arogozhnikov/einops/blob/master/tests/test_parsing.py
+
+import numpy as np
+import unittest
+from tinygrad import Tensor
+
+
+class test_rearrange_examples(unittest.TestCase):
+  def test1(self):
+    # transpose
+    x = Tensor(np.arange(10 * 20 * 30 * 40, dtype=np.int32).reshape([10, 20, 30, 40]))
+    y = x.rearrange("b c h w -> b h w c")
+    assert tuple(y.shape) == (10, 30, 40, 20)
+
+  def test2(self):
+    # view / reshape
+    x = Tensor(np.arange(10 * 20 * 30 * 40, dtype=np.int32).reshape([10, 20, 30, 40]))
+    y = x.rearrange("b c h w -> b (c h w)")
+    assert tuple(y.shape) == (10, 20 * 30 * 40)
+
+  def test3(self):
+    # depth-to-space
+    x = Tensor(np.arange(10 * 20 * 30 * 40, dtype=np.int32).reshape([10, 20, 30, 40]))
+    y = x.rearrange("b (c h1 w1) h w -> b c (h h1) (w w1)", h1=2, w1=2)
+    assert tuple(y.shape) == (10, 5, 30 * 2, 40 * 2)
+
+  def test4(self):
+    # space-to-depth
+    x = Tensor(np.arange(10 * 20 * 30 * 40, dtype=np.int32).reshape([10, 20, 30, 40]))
+    y = x.rearrange("b c (h h1) (w w1) -> b (h1 w1 c) h w", h1=2, w1=2)
+    assert tuple(y.shape) == (10, 20 * 4, 30 // 2, 40 // 2)
+
+  def test5(self):
+    # simple transposition
+    x = Tensor(np.arange(10 * 20 * 30 * 40, dtype=np.int32).reshape([10, 20, 30, 40]))
+    y = x.rearrange("b1 sound b2 letter -> b1 b2 sound letter")
+    assert tuple(y.shape) == (10, 30, 20, 40)
+
+  def test6(self):
+    # parsing parameters
+    x = Tensor(np.arange(10 * 20 * 30 * 40, dtype=np.int32).reshape([10, 20, 30, 40]))
+    t = x.rearrange("b c h w -> (b h w) c")
+    t = t[:, ::2]  # replacement for dot-product, just changes size of second axis
+    assert tuple(t.shape) == (10 * 30 * 40, 10)
+
+  def test7(self):
+    x = Tensor(np.arange(10 * 20 * 30 * 40, dtype=np.int32).reshape([10, 20, 30, 40]))
+    # split of embedding into groups
+    y1, y2 = x.rearrange("b (c g) h w -> g b c h w", g=2)
+    assert tuple(y1.shape) == (10, 10, 30, 40)
+    assert tuple(y2.shape) == (10, 10, 30, 40)
+
+  def test8(self):
+    x = Tensor(np.arange(10 * 20 * 1 * 1, dtype=np.int32).reshape([10, 20, 1, 1]))
+    # squeeze - unsqueeze
+    y = x.rearrange("b c () () -> b c")
+    assert tuple(y.shape) == (10, 20)
+    y = y.rearrange("b c -> c b () ()")
+    assert tuple(y.shape) == (20, 10, 1, 1)
+
+  def test9(self):
+    x = Tensor(np.arange(10 * 20 * 1 * 1, dtype=np.int32).reshape([10, 20, 1, 1]))
+    # squeeze - unsqueeze
+    y = x.rearrange("b c 1 1 -> b c")
+    assert tuple(y.shape) == (10, 20)
+    y = y.rearrange("b1 c -> c b1 1 1")
+    assert tuple(y.shape) == (20, 10, 1, 1)
+
+
+class test_rearrange_ops(unittest.TestCase):
+  def test_rearrange_errors(self):
+    x = Tensor.zeros([1, 1, 1, 1, 1])
+    x.rearrange("a b c d ... ->  a b c ... d")
+    bad_patterns = [
+      "a b c d (...) ->  a b c ... d",  # collapsed ellipsis on input
+      "a b (c d ... ->  a b c ... d",   # unbalanced brackets
+      "a b* c d ... ->  a b c ... d",   # not alphanumeric
+      "a b c d ->  a b c d -> a b c d", # two "->"
+      "a ... c ... ->  ... a ... c",    # two "..."
+      "a b c d e -> f b c d e",         # name mismatch
+    ]
+    for pattern in bad_patterns:
+      with self.assertRaises(AssertionError):
+        x.rearrange(pattern)
+
+    x.rearrange("... ->  (...)")
+    with self.assertRaises(AssertionError):
+      x.rearrange("(...) -> (...)")
+
+    y = Tensor.zeros([8, 1])
+    y.rearrange("(a1 a2 a3) b -> b a3 a2 a1", a1=2, a2=2)
+    with self.assertRaises(RuntimeError):
+      ## should fail as not enough dimensions specified
+      y.rearrange("(a1 a2 a3) b -> b a3 a2 a1", a1=2)
+    with self.assertRaises(ValueError):
+      ## should fail as 6 does not divide 8
+      y.rearrange("(a1 a2 a3) b -> b a3 a2 a1", a1=3, a2=2)
+    with self.assertRaises(AssertionError):
+      ## incorrect dimension provided for an axis that is only permuted
+      y.rearrange("(a1 a2 a3) b -> b a3 a2 a1", a1=2, a2=2, b=2)
+    with self.assertRaises(AssertionError):
+      ## unused axis provided
+      y.rearrange("(a b c) d -> a b c d", b=2, c=2, e=2)
+
+
+class test_rearrange_parsing(unittest.TestCase):
+  def test_elementary_axis_name(self):
+    for name in [
+      "a",
+      "b",
+      "h",
+      "dx",
+      "h1",
+      "zz",
+      "i9123",
+      "somelongname",
+      "Alex",
+      "camelCase",
+      "u_n_d_e_r_score",
+      "unreasonablyLongAxisName",
+    ]:
+      Tensor.ones((1,)).rearrange(f"{name} -> {name}")
+
+    for name in ["2b", "12", "_startWithUnderscore", "endWithUnderscore_", "_"]:
+      with self.assertRaises(AssertionError):
+        Tensor.ones((1,)).rearrange(f"{name} -> {name}")
+
+    with self.assertRaises(RuntimeError):
+      Tensor.ones((1,)).rearrange(" -> ")
+
+  def test_invalid_expressions(self):
+    # double ellipsis should raise an error
+    def _test_expression(expression: str):
+      Tensor.ones((2, 3, 4, 5, 6)).rearrange(f"{expression} -> {expression}")
+
+    _test_expression("... a b c d")
+    with self.assertRaises(AssertionError):
+      _test_expression("... a b c d ...")
+    with self.assertRaises(AssertionError):
+      _test_expression("... a b c (d ...)")
+    with self.assertRaises(AssertionError):
+      _test_expression("(... a) b c (d ...)")
+
+    # double/missing/enclosed parenthesis
+    Tensor.ones((2, 3, 4, 5, 6)).rearrange("a b c d ... -> (a) b c (d ...)")
+    with self.assertRaises(AssertionError):
+      _test_expression("(a)) b c (d ...)")
+    with self.assertRaises(AssertionError):
+      _test_expression("(a b c (d ...)")
+    with self.assertRaises(AssertionError):
+      _test_expression("(a) (()) b c (d ...)")
+    with self.assertRaises(AssertionError):
+      _test_expression("(a) ((b c) (d ...))")
+
+    # invalid identifiers
+    _test_expression("camelCase under_scored cApiTaLs ß ...")
+    with self.assertRaises(AssertionError):
+      _test_expression("1a")
+    with self.assertRaises(AssertionError):
+      _test_expression("_pre")
+    with self.assertRaises(AssertionError):
+      _test_expression("...pre")
+    with self.assertRaises(AssertionError):
+      _test_expression("pre...")
+
+
+if __name__ == "__main__":
+  unittest.main()

test/null/test_resnet.py

@@ -1,10 +1,6 @@
 import unittest
 from extra.models import resnet
-from tinygrad import dtypes
-from tinygrad.device import is_dtype_supported
 
-# pretrained weights contain num_batches_tracked as int64
-@unittest.skipUnless(is_dtype_supported(dtypes.int64), "need int64 support")
 class TestResnet(unittest.TestCase):
   def test_model_load(self):
     model = resnet.ResNet18()

test/null/test_schedule_cache.py

@@ -0,0 +1,41 @@
+import unittest
+from tinygrad import Tensor, Variable, Context
+from tinygrad.helpers import cpu_events
+from tinygrad.engine.schedule import schedule_cache
+
+def schedule_one():
+  Tensor([1]).schedule()
+
+class TestScheduleCache(unittest.TestCase):
+  def test_bound_variable_var_vals(self):
+    v = Variable('pos', 1, 100)
+    x = Tensor.ones(10).contiguous().realize()
+
+    t = x + Tensor(v.bind(42))
+    _, var_vals = t.schedule_with_vars()
+    self.assertEqual(var_vals, {'pos': 42})
+
+  def test_disable_schedule_cache(self):
+    schedule_cache.clear()
+
+    # test write
+    with Context(SCACHE=0): schedule_one()
+    self.assertEqual(len(schedule_cache), 0)
+    with Context(SCACHE=1):
+      schedule_one()
+      schedule_one()
+    self.assertEqual(len(schedule_cache), 1)
+
+    # test read
+    with Context(PROFILE=1):
+      cpu_events.clear()
+      with Context(SCACHE=0): schedule_one()
+      num_events_no_cache = len(cpu_events)
+
+      cpu_events.clear()
+      with Context(SCACHE=1): schedule_one()
+      num_events_cache = len(cpu_events)
+    self.assertLess(num_events_cache, num_events_no_cache)
+
+if __name__ == "__main__":
+  unittest.main()

test/null/test_symbolic_tensor.py

@@ -0,0 +1,101 @@
+import unittest
+from tinygrad import Variable
+from tinygrad.tensor import Tensor
+
+class TestSymbolic(unittest.TestCase):
+  def assert_tuple_equal(self, x, y):
+    for a,b in zip(x,y): self.assertFalse(a != b)
+
+  def test_cat_dim0_is_expanded(self):
+    i = Variable("i", 1, 5).bind(3)
+    j = Variable("j", 1, 5).bind(3)
+    k = Variable("k", 1, 5).bind(3)
+    t = Tensor.rand(5, 4)[:i].cat(Tensor.rand(5, 4)[:j], dim=0).cat(Tensor.rand(5, 4)[:k], dim=0)
+    self.assert_tuple_equal(t.shape, (i+j+k, 4))
+    t = Tensor.rand(5, 3)[:i].cat(Tensor.rand(5, 3)[:i], dim=0).cat(Tensor.rand(3, 3), dim=0)
+    self.assert_tuple_equal(t.shape, (2*i+3, 3))
+
+  def test_cat_dim1_strides(self):
+    i = Variable("i", 1, 5).bind(4)
+    j = Variable("j", 1, 5).bind(4)
+    k = Variable("k", 1, 5).bind(4)
+    t = Tensor.rand(3, 5)[:, :i].cat(Tensor.rand(3, 5)[:, :j], dim=1).cat(Tensor.rand(3, 5)[:, :k], dim=1)
+    self.assert_tuple_equal(t.shape, (3, i+j+k))
+
+class TestSymbolicVarVals(unittest.TestCase):
+  def assert_equal(self, x, y): self.assertFalse(x != y)
+
+  def test_shrink_unbind(self):
+    v = Variable("v", 1, 100)
+    bv = Variable("v", 1, 100).bind(2)
+    t = Tensor.rand(3, 4).shrink(((0,bv),(0,4)))
+    unbound_st, var_val = t.uop.unbind_all()
+    assert var_val == {v: 2}
+    t = Tensor.rand(3, 4).shrink(((bv, bv+1), (0, 4)))
+    unbound_st, var_val = t.uop.unbind_all()
+    assert var_val == {v: 2}
+
+class TestSymbolicReshape(unittest.TestCase):
+  def test_reshape(self):
+    a = Tensor.rand(5, 4)
+    b = Tensor.rand(5, 6)
+    for i in range(1, 6):
+      vi = Variable("i", 1, 5).bind(i)
+      ret = a[:vi]
+      ret = ret.reshape((vi, 4))
+      assert ret.shape == (vi, 4)
+      ret = b[:vi]
+      ret = ret.reshape((vi, 2, 3))
+      assert ret.shape == (vi, 2, 3)
+
+  def test_two_symbol_reshape(self):
+    t = Tensor.rand(5, 5)
+    for i in range(1, 6):
+      for j in range(1, 6):
+        vi = Variable("i", 1, 5).bind(i)
+        vj = Variable("j", 1, 5).bind(j)
+        ret = t[:vi, :vj]
+        ret = ret.reshape(vj, vi)
+        assert ret.shape == (vj, vi)
+        ret = ret.reshape(vi, vj)
+        assert ret.shape == (vi, vj)
+        ret = ret.reshape(1, vi*vj)
+        assert ret.shape == (1, vi*vj)
+
+class TestSymbolicExpand(unittest.TestCase):
+  def test_expand_into_symbols(self):
+    vi = Variable("i", 1, 5).bind(3)
+    vj = Variable("j", 1, 5).bind(3)
+    a = Tensor([[1], [2], [3]]).expand((3, vi))
+    assert a.shape == (3, vi)
+    a = a.reshape(3, vi, 1).expand((3, vi, vj))
+    assert a.shape == (3, vi, vj)
+
+  def test_plus_expands_constant(self):
+    a = Tensor.rand(3, 5)
+    for i in range(1, 6):
+      vi = Variable("i", 1, 5).bind(i)
+      ret = a[:, :vi]
+      ret = ret + 1
+      self.assertTupleEqual(ret.shape, (3, vi))
+
+  def test_pad_then_expand_into_symbols(self):
+    vi = Variable("i", 1, 10).bind(3)
+    a = Tensor(1).unsqueeze(0).pad((0, 24)).unsqueeze(0).expand((vi, 25))
+    self.assertEqual(a.shape, (vi, 25))
+    self.assertEqual(a.reshape(25*vi).shape, (vi*25,))
+    self.assertEqual(a.reshape(vi*25).shape, (vi*25,))
+
+class TestSymbolicShrink(unittest.TestCase):
+  def test_shrink_symbols_simple(self):
+    vi = Variable("i", 1, 5)
+    t = Tensor.rand(5, 5).shrink(((0, 5),(0,vi)))
+    assert t.shape == (5, vi)
+
+  def test_shrink_symbols(self):
+    vi = Variable("i", 1, 5)
+    t = Tensor.rand(3, 5).shrink(((0, 2), (vi, vi+1)))
+    assert t.shape == (2, 1)
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_tensor.py

@@ -0,0 +1,166 @@
+# tensor tests that pass on NULL backend (no copyout needed)
+import numpy as np
+import unittest
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.device import is_dtype_supported
+from tinygrad.uop.ops import Ops, UOp
+from tinygrad.renderer.ptx import PTXRenderer
+from tinygrad.renderer.nir import NIRRenderer
+from tinygrad.engine.realize import get_program
+from tinygrad.dtype import DType
+
+x_init = np.random.randn(1,3).astype(np.float32)
+W_init = np.random.randn(3,3).astype(np.float32)
+m_init = np.random.randn(1,3).astype(np.float32)
+
+class TestTrainMode(unittest.TestCase):
+  def test_train_mode(self):
+    assert not Tensor.training
+    @Tensor.train()
+    def f():
+      assert Tensor.training
+    f()
+    assert not Tensor.training
+
+class TestInferenceMode(unittest.TestCase):
+  def test_inference(self):
+    x = Tensor(x_init, requires_grad=True)
+    m = Tensor(m_init, requires_grad=True)
+    W = Tensor(W_init, requires_grad=True)
+    tmp = x.mul(m)
+    mm = tmp.matmul(W)
+    out = mm.relu()
+    out = out.sum()
+    #out.backward()
+    assert x.grad is None
+    assert m.grad is None
+    assert tmp.grad is None
+    assert mm.grad is None
+    assert W.grad is None
+    assert W.requires_grad
+
+  def test_no_grad_mode_context_manager(self):
+    x = Tensor(x_init, requires_grad=True)
+    m = Tensor(m_init, requires_grad=True)
+    W = Tensor(W_init, requires_grad=True)
+    def f(x, m, W):
+      tmp = x.mul(m)
+      mm = tmp.matmul(W)
+      out = mm.relu()
+      out = out.sum()
+      #out.backward()
+      assert x.grad is None
+      assert m.grad is None
+      assert tmp.grad is None
+      assert mm.grad is None
+      assert W.grad is None
+    f(x, m, W)
+
+class TestIdxUpcast(unittest.TestCase):
+  def _find_op(self, ast: UOp, op: Ops):
+    if ast.op is op: return ast
+    for src in ast.src:
+      if (ret:=self._find_op(src, op)) is not None: return ret
+  def _schedule_render(self, a: Tensor):
+    schedule, _ = a.schedule_with_vars()
+    for s in schedule:
+      if s.ast.op is Ops.SINK:
+        renderer = Device[s.bufs[0].device].renderer
+        prg = get_program(s.ast, renderer)
+        return prg.uops
+
+  def _assert(self, dtype: DType, a: Tensor):
+    uops = self._schedule_render(a)
+    # Assert the dtype of the INDEX value, This will need be updated if UOp spec changes
+    store = next(uop for uop in uops if uop.op is Ops.STORE)
+    assert store.op is Ops.STORE
+    idx = self._find_op(store, Ops.INDEX)
+    # PTX and NIR turn Ops.INDEX into pointer arithmetic earlier than cstyle, plus it's already cast to int64
+    if not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)):
+      assert idx.op is Ops.INDEX
+      idx_val = idx.src[1]
+      assert idx_val.dtype is dtype
+
+  # use expand to generate kernel that uses large idx
+  def do_op_then_assert(self, dtype: DType, dim1, dim2, dim3):
+    self._assert(dtype, Tensor.empty(dim1, dim2, 1).expand(-1, -1, dim3).contiguous())
+
+  @unittest.skipUnless(is_dtype_supported(dtypes.long), "int64 is supported")
+  def test_overflow(self):
+    # 2**11, 2**11, 2**11 -> 2**33 will overflow when indexed
+    self.do_op_then_assert(dtypes.long, 2048, 2048, 2048)
+
+  @unittest.skipUnless(is_dtype_supported(dtypes.long), "int64 is supported")
+  def test_overflow_sym(self):
+    self.do_op_then_assert(dtypes.long, 2048, 2048, UOp.variable("dim3", 1, 2048).bind(32))
+
+  def test_regular(self):
+    self.do_op_then_assert(dtypes.int, 64, 64, 64)
+
+  def test_regular_sym(self):
+    self.do_op_then_assert(dtypes.int, 2048, 2048, UOp.variable("dim3", 1, 64).bind(32))
+
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)), "PTX and NIR always converts Ops.INDEX to int64")
+  def test_symfold(self):
+    # This would cause an overflow, but after sym fold it's within int32
+    a = Tensor.arange(65535)
+    uops = self._schedule_render(a)
+    assert all(uop.dtype is not dtypes.long for uop in uops)
+
+  def test_arange_raise_overflow(self):
+    with self.assertRaises(ValueError):
+      self._schedule_render(Tensor.arange(2**33, dtype=dtypes.int))
+
+  @unittest.skipIf(is_dtype_supported(dtypes.long), "int64 is supported")
+  def test_int64_unsupported_overflow_sym(self):
+    with self.assertRaises(KeyError):
+      self.do_op_then_assert(dtypes.long, 2048, 2048, UOp.variable("dim3", 1, 2048).bind(32))
+
+  @unittest.skipIf(is_dtype_supported(dtypes.long), "int64 is supported")
+  @unittest.expectedFailure  # bug in gpu dims limiting
+  def test_int64_unsupported_overflow(self):
+    with self.assertRaises(KeyError):
+      self.do_op_then_assert(dtypes.long, 2048, 2048, 2048)
+
+  @unittest.skip("This is kept for reference, it requires large memory to run")
+  def test_overflow_kernel_run(self):
+    # This creates a total of 2**31+10 elements, requiring at least 2147 MB memory to run
+    # Modified example from issue 3271
+    a = Tensor.empty(2**11, 2**11, 1, dtype=dtypes.int8).permute((2, 0, 1)).expand((2**9+10, -1, -1)).contiguous()
+    a.realize()
+
+class TestTensorUnique(unittest.TestCase):
+  def test_empty_bufs_unique(self):
+    a = Tensor.empty(10, 10).contiguous()
+    b = Tensor.empty(10, 10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_zeros_bufs_unique_sep(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(a)
+    b = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_zeros_bufs_unique(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    b = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_eye_bufs_unique(self):
+    a = Tensor.eye(10).contiguous()
+    b = Tensor.eye(10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_times_2_not_unique(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    b = a * 2
+    c = a * 2
+    Tensor.realize(b,c)
+    self.assertIs(b.uop.buffer, c.uop.buffer)
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_tensor_io.py

@@ -0,0 +1,24 @@
+import unittest
+from tinygrad import Tensor, dtypes
+from tinygrad.nn.state import TensorIO
+
+class TestTensorIO(unittest.TestCase):
+  def test_create(self):
+    with self.assertRaises(ValueError):
+      TensorIO(Tensor(b"Hello World").reshape(1, -1))
+    with self.assertRaises(ValueError):
+      TensorIO(Tensor([], dtype=dtypes.int64).reshape(1, -1))
+
+  def test_seek(self):
+    t = Tensor(b"Hello World!")
+    fobj = TensorIO(t)
+    self.assertEqual(fobj.tell(), 0)
+    self.assertEqual(fobj.seek(1), 1)
+    self.assertEqual(fobj.seek(-2, 2), len(t) - 2)
+    self.assertEqual(fobj.seek(1, 1), len(t) - 1)
+    self.assertEqual(fobj.seek(10, 1), len(t))
+    self.assertEqual(fobj.seek(10, 2), len(t))
+    self.assertEqual(fobj.seek(-10, 0), 0)
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_tinyfs.py

@@ -5,18 +5,22 @@ class TestLoadStore(unittest.TestCase):
   def test_load_shape(self):
     t = Tensor(bytes(16)).fs_load(1024)
     assert t.shape == (1024,), t.shape
+    t.schedule()
 
   def test_store_shape(self):
     t = Tensor.zeros(1024).fs_store()
     assert t.shape == (16,), t.shape
+    t.schedule()
 
   def test_load_large_shape(self):
     t = Tensor(bytes(16)).fs_load(10_000_000)
     assert t.shape == (10_000_000,), t.shape
+    t.schedule()
 
   def test_store_large_shape(self):
     t = Tensor.zeros(10_000_000).fs_store()
     assert t.shape == (16,), t.shape
+    t.schedule()
 
 if __name__ == "__main__":
   unittest.main()

test/null/test_uops.py

@@ -0,0 +1,334 @@
+# uops tests that pass on NULL backend (no copyout needed)
+import unittest
+import numpy as np
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import Timing, Context
+from tinygrad.dtype import dtypes, ConstFloat  # noqa: F401
+from tinygrad.device import Device
+from tinygrad.uop.ops import Ops, UOp, UPat, KernelInfo, exec_alu
+from tinygrad.uop.spec import shared_spec
+from tinygrad.uop.symbolic import sym
+from test.helpers import get_uops
+
+def to_uops_list(u:list[UOp], ren=None) -> list[UOp]:
+  sink = UOp.group(*u)
+  for r in sink.ranges: sink = sink.end(r)
+  ret = get_uops(sink.sink(arg=KernelInfo(opts_to_apply=())), ren)
+  assert ret[-1].op is Ops.SINK
+  return ret
+
+class TestSafeCast(unittest.TestCase):
+  def test_cast_folds(self):
+    a = UOp.variable("a", 1, 10, dtype=dtypes.int32)
+    self.assertEqual(a.cast(dtypes.int64).cast(dtypes.int32).simplify(), a)
+    self.assertEqual(a.cast(dtypes.double).cast(dtypes.int32).simplify(), a)
+    a = UOp.variable("a", 1, 10, dtype=dtypes.uint8)
+    self.assertEqual(a.cast(dtypes.int64).cast(dtypes.uint8).simplify(), a)
+    self.assertEqual(a.cast(dtypes.uint32).cast(dtypes.uint8).simplify(), a)
+
+  def test_remove_intermediate_cast(self):
+    a = UOp.variable("a", 0., 100., dtype=dtypes.half)
+    self.assertEqual(a.cast(dtypes.double).cast(dtypes.float).simplify(), a.cast(dtypes.float))
+    a = UOp.variable("a", 1, 10, dtype=dtypes.int32)
+    # TODO: double preserves certain int dtypes
+    self.assertEqual(a.cast(dtypes.double).cast(dtypes.float).simplify(), a.cast(dtypes.float))
+    self.assertEqual(a.cast(dtypes.int64).cast(dtypes.int16).simplify(), a.cast(dtypes.int16))
+    a = UOp.variable("a", 1, 10, dtype=dtypes.uint8)
+    self.assertEqual(a.cast(dtypes.int64).cast(dtypes.int32).simplify(), a.cast(dtypes.int32))
+
+  def test_safe_cast_using_bounds(self):
+    a = UOp.variable("a", 1, 10, dtype=dtypes.uint64)
+    self.assertEqual(a.cast(dtypes.int16).cast(dtypes.int).simplify(), a.cast(dtypes.int))
+    a = UOp.variable("a", -10, 10, dtype=dtypes.int32)
+    self.assertEqual(a.cast(dtypes.int8).cast(dtypes.int64).simplify(), a.cast(dtypes.int64))
+    self.assertEqual(a.cast(dtypes.int8).cast(dtypes.float).simplify(), a.cast(dtypes.float))
+
+class TestExecALU(unittest.TestCase):
+  def test_sqrt(self):
+    self.assertEqual(exec_alu(Ops.SQRT, dtypes.float, (0.0,)), 0.0)
+
+  def test_div(self):
+    self.assertEqual(exec_alu(Ops.IDIV, dtypes.int8, (8, 2)), 4)
+    self.assertEqual(exec_alu(Ops.IDIV, dtypes.int8, (7, 3)), 2)
+    self.assertEqual(exec_alu(Ops.IDIV, dtypes.int8, (7, -3)), -2)
+    self.assertEqual(exec_alu(Ops.IDIV, dtypes.int8, (-50, 6)), -8)
+
+    np.testing.assert_allclose(exec_alu(Ops.MUL, dtypes.float32, (7.0, exec_alu(Ops.RECIPROCAL, dtypes.float32, (3.0,)))), 2+(1.0/3.0))
+    np.testing.assert_allclose(exec_alu(Ops.MUL, dtypes.float32, (7.0, exec_alu(Ops.RECIPROCAL, dtypes.float32, (-3.0,)))), -2-(1.0/3.0))
+
+  def test_recip(self):
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (8,)), 1/8)
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (7,)), 1/7)
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (-3,)), 1/-3)
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (-50,)), 1/-50)
+
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, ((32+521+3),)), 1/(32+521+3))
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, ((34**2),)), 1/(34**2))
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (10,)), 1/10)
+
+  def test_bool_cmplt(self):
+    self.assertEqual(exec_alu(Ops.CMPLT, dtypes.bool, (False, False)), False)
+    self.assertEqual(exec_alu(Ops.CMPLT, dtypes.bool, (False, True)), True)
+    self.assertEqual(exec_alu(Ops.CMPLT, dtypes.bool, (True, False)), False)
+    self.assertEqual(exec_alu(Ops.CMPLT, dtypes.bool, (True, True)), False)
+
+  def test_bool_cmpne(self):
+    self.assertEqual(exec_alu(Ops.CMPNE, dtypes.bool, (False, False)), False)
+    self.assertEqual(exec_alu(Ops.CMPNE, dtypes.bool, (False, True)), True)
+    self.assertEqual(exec_alu(Ops.CMPNE, dtypes.bool, (True, False)), True)
+    self.assertEqual(exec_alu(Ops.CMPNE, dtypes.bool, (True, True)), False)
+
+  def test_bool_where(self):
+    self.assertEqual(exec_alu(Ops.WHERE, dtypes.bool, (False, False, False)), False)
+    self.assertEqual(exec_alu(Ops.WHERE, dtypes.int, (False, 2, 4)), 4)
+    np.testing.assert_allclose(exec_alu(Ops.WHERE, dtypes.float, (False, 2.2, 4.5)), 4.5)
+
+  def test_overflow(self):
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.uint8, (250, 250)), 244)
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.uint8, (256, 0)), 0)
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.uint8, (0, -1)), 255)
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.uint8, (0, -1000)), 24)
+
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.int8, (127, 0)), 127)
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.int8, (-128, 0)), -128)
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.int8, (-100, -100)), 56)
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.int8, (-1000, -0)), 24)
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.int8, (-130, -0)), 126)
+
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.int8, (1, 1)), 2)
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.int8, (-128, 0)), -128)
+
+    # test no truncate
+    self.assertEqual(exec_alu(Ops.ADD, dtypes.uint8, (250, 250), truncate_output=False), 500)
+
+class TestConstantFolding(unittest.TestCase):
+  def test_cast_const(self):
+    t = Tensor(1, dtype=dtypes.float).cast(dtypes.int)
+    si = t.schedule()
+    assert len(si) == 0
+
+class TestGatedStoreRewrite(unittest.TestCase):
+  def test_tiny_gate_store(self):
+    gmem = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
+    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
+    gate = gidx0<UOp.const(dtypes.int, 1)
+    idx = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem, (gidx0 * UOp.const(dtypes.int, 2)).valid(gate)))
+    val = UOp.const(dtypes.float, 42.0)
+    store = UOp(Ops.STORE, dtypes.void, (idx, val))
+    uops = to_uops_list([store])
+    if_uop = next(u for u in uops if u.op is Ops.IF)
+    endif = next(u for u in uops if u.op is Ops.ENDIF)
+    assert endif.src[0] is if_uop
+    gated_uops = tuple(uops[uops.index(if_uop)+1:uops.index(endif)])
+    self.assertEqual(len(gated_uops), 1)
+    self.assertIs(gated_uops[-1].op, Ops.STORE)
+
+  def test_gate_some_stores(self):
+    gmem0 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
+    gmem1 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 1)
+    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
+    idx = gidx0 * UOp.const(dtypes.int, 2)
+    idx0 = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem0, idx.valid(gidx0<UOp.const(dtypes.int, 1))))
+    idx1 = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem1, idx))
+    val = UOp.const(dtypes.float, 42.0)
+    stores = [UOp.store(idx0, val), UOp.store(idx1, val)]
+    uops = to_uops_list(stores)
+    if_uop = next(u for u in uops if u.op is Ops.IF)
+    endif = next(u for u in uops if u.op is Ops.ENDIF)
+    assert endif.src[0] is if_uop
+    gated_uops = tuple(uops[uops.index(if_uop)+1:uops.index(endif)])
+    self.assertEqual(len(gated_uops), 1)
+    self.assertIs(gated_uops[-1].op, Ops.STORE)
+
+  # scaled down version of TestLinearizerDumb.test_unmerged_ifs
+  @unittest.skip("we don't merge ifs anymore")
+  def test_merge_ifs_alt(self):
+    gmem0 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
+    gmem1 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 1)
+    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
+    idx = gidx0*UOp.const(dtypes.int, 2)
+    gate = gidx0<UOp.const(dtypes.int, 1)
+    idx0 = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem0, idx, gate))
+    idx1 = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem1, idx, gate))
+    val = UOp.const(dtypes.float, 42.0)
+    stores = [UOp.store(idx0, val), UOp.store(idx1, val)]
+    uops = to_uops_list(stores)
+    ifs = [u for u in uops if u.op is Ops.IF]
+    endifs = [u for u in uops if u.op is Ops.ENDIF]
+    self.assertEqual(len(ifs), 1)
+    self.assertEqual(len(endifs), 1)
+    gated_uops = tuple(uops[uops.index(ifs[0])+1:uops.index(endifs[0])])
+    self.assertEqual(len(gated_uops), 2)
+    for x in gated_uops: self.assertIs(x.op, Ops.STORE)
+
+@unittest.skipIf(Device.DEFAULT == "METAL", "compiler bug")
+@unittest.skipUnless(Ops.SHR in Device[Device.DEFAULT].renderer.code_for_op, "fast_idiv requires SHR")
+class TestFastIdiv(unittest.TestCase):
+  def test_division_power_of_two(self):
+    for dt in (dtypes.int32, dtypes.uint32):
+      g = UOp(Ops.PARAM, dt.ptr(), (), 0)
+      c = UOp.const(dt, 2)
+      l = g.index(c)
+      a = UOp(Ops.IDIV, dt, (l, c))
+      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
+      Device[Device.DEFAULT].renderer.render(uops)
+      ops = [x.op for x in uops]
+      self.assertIn(Ops.SHR, ops, f"For dtype={dt} divison by power of two did not simplify to shift")
+      self.assertNotIn(Ops.IDIV, ops, f"For dtype={dt} divison by power of two did not simplify to shift")
+
+  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "WEBGPU doesn't support long")
+  def test_fast_idiv_and_mod(self):
+    g = UOp(Ops.PARAM, dtypes.uint32.ptr(), (), 0)
+    c = UOp.const(dtypes.uint, 3)
+    l = g.index(c)
+    a = UOp(Ops.IDIV, dtypes.uint, (l, c))
+    uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
+    Device[Device.DEFAULT].renderer.render(uops)
+    ops = [x.op for x in uops]
+    self.assertIn(Ops.SHR, ops)
+    self.assertNotIn(Ops.IDIV, ops)
+
+    b = UOp(Ops.MOD, dtypes.uint, (l, c))
+    uops = to_uops_list([b], ren=Device[Device.DEFAULT].renderer)
+    Device[Device.DEFAULT].renderer.render(uops)
+    ops = [x.op for x in uops]
+    self.assertIn(Ops.SHR, ops)
+    self.assertNotIn(Ops.MOD, ops)
+
+  def test_fast_idiv_remove_powers_of_two(self):
+    ridx = UOp.range(2**20, 0)
+    uops = to_uops_list([ridx//(7*64)], ren=Device[Device.DEFAULT].renderer)
+    ops = [x.op for x in uops]
+    # this requires shifting out the powers of two before doing fast_idiv
+    # (((ridx0>>6)*18725)>>17) instead of (int)((((long)(ridx0)*1198373)>>29))
+    self.assertNotIn(Ops.CAST, ops)
+
+  @unittest.expectedFailure
+  def test_fast_idiv_overflow(self):
+    # This will be possible with a slightly different method for fast_idiv
+    g = UOp(Ops.PARAM, dtypes.uint32.ptr(), (), 0)
+    c = UOp.const(dtypes.uint, 7)
+    l = UOp(Ops.LOAD, dtypes.uint, (g.index(c),))
+    a = UOp(Ops.IDIV, dtypes.uint, (l, c))
+    uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
+    Device[Device.DEFAULT].renderer.render(uops)
+    ops = [x.op for x in uops]
+    self.assertIn(Ops.SHR, ops)
+    self.assertNotIn(Ops.IDIV, ops)
+
+  def test_disable_fast_idiv(self):
+    g = UOp(Ops.PARAM, dtypes.uint32.ptr(), (), 0)
+    c = UOp.const(dtypes.uint, 3)
+    l = g.index(c)
+    a = UOp(Ops.IDIV, dtypes.uint, (l, c))
+    with Context(DISABLE_FAST_IDIV=1):
+      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
+    ops = [x.op for x in uops]
+    self.assertNotIn(Ops.SHR, ops)
+    self.assertIn(Ops.IDIV, ops)
+
+class TestUOpMethod(unittest.TestCase):
+  @unittest.skip("uops lt no longer ordered")
+  def test_compare_alu_same_src_different_arg(self):
+    a = UOp.const(dtypes.float, 2.0)
+    b = UOp.const(dtypes.float, 3.0)
+
+    add = UOp(Ops.ADD, dtypes.float, (a, b))
+    mul = UOp(Ops.MUL, dtypes.float, (a, b))
+    assert (add < mul) or (mul < add), "add and mul with same src should have an order"
+
+  def test_uop_variables(self):
+    a = UOp.variable("a", 1, 10)
+    uop_var = Tensor(a.bind(1))
+    st_var = Tensor.empty((2, 10))[:, :a.bind(1)]
+    _, var_vals = (uop_var+st_var).schedule_with_vars()
+    self.assertEqual(len(var_vals), 1)
+    self.assertEqual(list(var_vals)[0], a.expr)
+
+  def test_const_factor(self):
+    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 8),), 'gidx0')
+    self.assertEqual(UOp.const(dtypes.int, 17).const_factor(), 17)
+    self.assertEqual(gidx0.const_factor(), 1)
+    self.assertEqual((gidx0*3).const_factor(), 3)
+    self.assertEqual((gidx0*3+6).const_factor(), 3)
+    self.assertEqual((gidx0*3+1).const_factor(), 1)
+
+  def test_replace(self):
+    x = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
+    self.assertIs(x.replace(arg=None).arg, None)
+    with self.assertRaises(AssertionError): x.replace(field="a")
+
+  def test_const_zero_neg_zero_different(self):
+    # -0.0 and 0.0 must be different UOps (for IEEE754 correctness, e.g. 1/-0.0 = -inf)
+    pos_zero = UOp.const(dtypes.float, 0.0)
+    neg_zero = UOp.const(dtypes.float, -0.0)
+    self.assertIsNot(pos_zero, neg_zero)
+    self.assertNotEqual(hash(pos_zero.arg), hash(neg_zero.arg))
+
+  def test_const_nan_same(self):
+    # nan constants should be deduplicated
+    nan1 = UOp.const(dtypes.float, float('nan'))
+    nan2 = UOp.const(dtypes.float, float('nan'))
+    self.assertIs(nan1, nan2)
+
+class TestUOpStr(unittest.TestCase):
+  def test_uop_str(self):
+    a = UOp.const(dtypes.float, 2.0) + UOp.const(dtypes.float, 3.0)
+    for _ in range(20): a = a + a
+    assert len(str(a)) < 10_000, "exponential string growth"
+    assert str(eval(str(a))) == str(a)
+
+  def test_vectorized_str(self):
+    vec = UOp(Ops.VECTORIZE, dtypes.int.vec(4), tuple(UOp.const(dtypes.int, x) for x in range(4)))
+    assert str(eval(str(vec))) == str(vec)
+
+  def test_device_arg(self):
+    device = UOp(Ops.DEVICE, arg="CL")
+    assert str(eval(str(device))) == str(device)
+
+  def test_reduceop_arg(self):
+    sum_uop = Tensor.empty(32, 32).sum().uop
+    assert str(eval(str(sum_uop))) == str(sum_uop)
+
+class TestUPatHelpers(unittest.TestCase):
+  def test_location(self):
+    self.assertEqual(sym.patterns[-1][0].location[0].replace("\\", "/").split("/")[-1], "symbolic.py")
+    self.assertEqual(shared_spec.patterns[0][0].location[0].replace("\\", "/").split("/")[-1], "spec.py")
+    test_upat = UPat(Ops.CONST, dtypes.bool)
+    self.assertEqual(test_upat.location[0].replace("\\", "/").split("/")[-1], __file__.replace("\\", "/").split("/")[-1])
+    test_upat_named = test_upat.named("test_name")
+    self.assertEqual(test_upat.location[0], test_upat_named.location[0])
+    self.assertNotEqual(test_upat.location[1], test_upat_named.location[1])
+
+class TestUopsObject(unittest.TestCase):
+  def test_timing(self):
+    with Timing("create 10k uops:"): ret = [UOp(Ops.CONST, dtypes.int, arg=10000000+i) for i in range(10000)]
+    assert len(ret) == 10000
+
+  def test_nested(self):
+    a = UOp.new_buffer(Device.DEFAULT, 1, dtypes.char)
+    for _ in range(10_000): a = a+a
+    self.assertEqual(a.device, Device.DEFAULT)
+
+class TestUOpRender(unittest.TestCase):
+  def test_render_vectorize_empty(self):
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(0), src=())
+    self.assertEqual(u.render(simplify=False), "{}")
+  def test_render_vectorize_empty_simplified(self):
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(0), src=())
+    self.assertEqual(u.render(), "{}")
+  def test_render_vectorize_same(self):
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0)))
+    self.assertEqual(u.render(simplify=False), "{0, ...}")
+  def test_render_vectorize_different(self):
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 2)))
+    self.assertEqual(u.render(simplify=False), "{0,1,2}")
+  def test_render_vectorize_same_simplified(self):
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0)))
+    self.assertEqual(u.render(), "0")
+  def test_render_vectorize_different_simplified(self):
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 2)))
+    self.assertEqual(u.render(), "(0, 1, 2)")
+
+if __name__ == '__main__':
+  unittest.main()

test/null/test_viz.py

@@ -353,8 +353,8 @@ def load_profile(lst:list[ProfileEvent]) -> dict:
     event_type, event_count = u("<BI")
     if event_type == 0:
       for _ in range(event_count):
-        name, ref, key, st, dur, _ = u("<IIIIfI")
-        v["events"].append({"name":strings[name], "ref":option(ref), "key":option(key), "st":st, "dur":dur})
+        name, ref, key, st, dur, fmt = u("<IIIIfI")
+        v["events"].append({"name":strings[name], "ref":option(ref), "key":option(key), "st":st, "dur":dur, "fmt":strings[fmt]})
     else:
       v["peak"] = u("<Q")[0]
       for _ in range(event_count):
@@ -365,8 +365,8 @@ def load_profile(lst:list[ProfileEvent]) -> dict:
 
 class TestVizProfiler(BaseTestViz):
   def test_node(self):
-    prof = [ProfileRangeEvent(device='NV', name='E_2', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=False),
-            ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100))]
+    prof = [ProfileRangeEvent(device='NV', name='E_2', st=decimal.Decimal(1000), en=decimal.Decimal(1010)),
+            ProfileDeviceEvent(device='NV', tdiff=decimal.Decimal(-1000))]
 
     j = load_profile(prof)
 
@@ -379,28 +379,28 @@ class TestVizProfiler(BaseTestViz):
     assert event['ref'] is None
 
   def test_copy_node(self):
-    prof = [ProfileRangeEvent(device='NV', name='COPYxx', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=True),
-            ProfileRangeEvent(device='NV:2', name='COPYxx', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=True),
-            ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100)),
-            ProfileDeviceEvent(device='NV:2', comp_tdiff=decimal.Decimal(-800), copy_tdiff=decimal.Decimal(-80))]
+    prof = [ProfileRangeEvent(device='NV:SDMA:0', name='COPYxx', st=decimal.Decimal(1000), en=decimal.Decimal(1010)),
+            ProfileRangeEvent(device='NV:2:SDMA:0', name='COPYxx', st=decimal.Decimal(1000), en=decimal.Decimal(1010)),
+            ProfileDeviceEvent(device='NV:SDMA:0', tdiff=decimal.Decimal(-100)),
+            ProfileDeviceEvent(device='NV:2:SDMA:0', tdiff=decimal.Decimal(-80))]
 
     j = load_profile(prof)
 
-    event = j['layout']['NV']['events'][0]
+    event = j['layout']['NV:SDMA:0']['events'][0]
     self.assertEqual(event['name'], 'COPYxx')
     self.assertEqual(event['st'], 0)   # first event
     self.assertEqual(event['dur'], 10)
 
-    event2 = j['layout']['NV:2']['events'][0]
+    event2 = j['layout']['NV:2:SDMA:0']['events'][0]
     self.assertEqual(event2['st'], 20) # second event, diff clock
 
     self.assertEqual(j["dur"], (event2["st"]+event2["dur"])-event["st"])
 
   def test_graph(self):
-    prof = [ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100)),
-            ProfileDeviceEvent(device='NV:1', comp_tdiff=decimal.Decimal(-500), copy_tdiff=decimal.Decimal(-50)),
-            ProfileGraphEvent(ents=[ProfileGraphEntry(device='NV', name='E_25_4n2', st_id=0, en_id=1, is_copy=False),
-                                    ProfileGraphEntry(device='NV:1', name='NV -> NV:1', st_id=2, en_id=3, is_copy=True)],
+    prof = [ProfileDeviceEvent(device='NV', tdiff=decimal.Decimal(-1000)),
+            ProfileDeviceEvent(device='NV:1:SDMA:0', tdiff=decimal.Decimal(-50)),
+            ProfileGraphEvent(ents=[ProfileGraphEntry(device='NV', name='E_25_4n2', st_id=0, en_id=1),
+                                    ProfileGraphEntry(device='NV:1:SDMA:0', name='NV -> NV:1', st_id=2, en_id=3)],
                               deps=[[], [0]],
                               sigs=[decimal.Decimal(1000), decimal.Decimal(1002), decimal.Decimal(1004), decimal.Decimal(1008)])]
 
@@ -409,22 +409,20 @@ class TestVizProfiler(BaseTestViz):
     tracks = list(j['layout'])
     self.assertEqual(tracks[0], 'NV')
     self.assertEqual(tracks[1], 'NV Graph')
-    self.assertEqual(tracks[2], 'NV:1')
+    self.assertEqual(tracks[2], 'NV:1:SDMA:0')
 
     nv_events = j['layout']['NV']['events']
     self.assertEqual(nv_events[0]['name'], 'E_25_4n2')
     self.assertEqual(nv_events[0]['st'], 0)
     self.assertEqual(nv_events[0]['dur'], 2)
-    #self.assertEqual(j['devEvents'][6]['pid'], j['devEvents'][0]['pid'])
 
-    nv1_events = j['layout']['NV:1']['events']
-    self.assertEqual(nv1_events[0]['name'], 'NV -> NV:1')
-    self.assertEqual(nv1_events[0]['st'], 954)
-    #self.assertEqual(j['devEvents'][7]['pid'], j['devEvents'][3]['pid'])
+    sdma_events = j['layout']['NV:1:SDMA:0']['events']
+    self.assertEqual(sdma_events[0]['name'], 'NV -> NV:1')
+    self.assertEqual(sdma_events[0]['st'], 954)
 
     graph_events = j['layout']['NV Graph']['events']
     self.assertEqual(graph_events[0]['st'], nv_events[0]['st'])
-    self.assertEqual(graph_events[0]['st']+graph_events[0]['dur'], nv1_events[0]['st']+nv1_events[0]['dur'])
+    self.assertEqual(graph_events[0]['st']+graph_events[0]['dur'], sdma_events[0]['st']+sdma_events[0]['dur'])
 
   def test_bytes_per_kernel(self):
     step = 10
@@ -465,11 +463,11 @@ class TestVizProfiler(BaseTestViz):
 
   def test_layout_order(self):
     def fn(): return
-    for dname in ["TINY", "USER", "TEST:1 N1", "TEST:2 N1", "TEST:1 N2"]:
+    for dname in ["TINY", "USER", "TEST:1 N1", "TEST:2 N1", "TEST:1 N2", "TEST:1:ENGINE:0", "TEST:1"]:
       with cpu_profile("fn", dname): fn()
     layout = list(load_profile(cpu_events)["layout"])
     self.assertListEqual(layout[:2], ["USER","TINY"])
-    self.assertListEqual(layout[2:], ["TEST:1 N1","TEST:1 N2", "TEST:2 N1"])
+    self.assertListEqual(layout[2:], ["TEST:1", "TEST:1:ENGINE:0", "TEST:1 N1","TEST:1 N2", "TEST:2 N1"])
 
 def _alloc(b:int):
   a = Tensor.empty(b, device="NULL", dtype=dtypes.char)

test/null/test_winograd.py

@@ -0,0 +1,59 @@
+import unittest, sys
+from tinygrad import Tensor, GlobalCounters, dtypes, Context
+from tinygrad.helpers import CI, Profiling, WINO
+
+@unittest.skipIf(sys.platform.startswith("win"), "flaky on Windows")
+class TestWinograd(unittest.TestCase):
+  def setUp(self):
+    self.old = WINO.value
+    WINO.value = 1
+  def tearDown(self):
+    WINO.value = self.old
+
+  def test_profile(self):
+    x,w = Tensor.rand(1,4,9,9).realize(), Tensor.rand(4,4,3,3).realize()
+    with Profiling(enabled=not CI, sort='time'):
+      Tensor.conv2d(x,w).realize()
+
+  def test_forward_kernels(self):
+    x,w = Tensor.rand(1,4,9,9).realize(), Tensor.rand(4,4,3,3).realize()
+    out = Tensor.conv2d(x,w)
+    self.assertEqual(len(out.schedule()), 2)
+
+  def test_backward_kernels(self):
+    x,w = Tensor.empty(1,4,9,9,requires_grad=True).realize(), Tensor.empty(4,4,3,3,requires_grad=True).realize()
+    out = Tensor.conv2d(x,w, padding=1)
+    out.mean().backward()
+    backward_schedule = Tensor.schedule(x.grad, w.grad)
+    self.assertEqual(len(backward_schedule), 4)
+
+  def test_counters(self):
+    IC, OC, X, Y = 4,4,9,9
+    x,w = Tensor.rand(1,IC,Y,X).realize(), Tensor.rand(OC,IC,3,3).realize()
+    GlobalCounters.reset()
+    with Context(WINO=1):
+      Tensor.conv2d(x,w).realize()
+    ops_wino, mem_wino = GlobalCounters.global_ops, GlobalCounters.global_mem
+    GlobalCounters.reset()
+    with Context(WINO=0):
+      Tensor.conv2d(x,w).realize()
+    ops_normal, mem_normal = GlobalCounters.global_ops, GlobalCounters.global_mem
+
+    ops_ratio, mem_ratio = ops_wino/ops_normal, mem_wino/mem_normal
+    print(f"ops: normal {ops_normal:9d} wino {ops_wino:9d} ratio {ops_ratio:.2f}")
+    print(f"mem: normal {mem_normal:9d} wino {mem_wino:9d} ratio {mem_ratio:.2f}")
+
+    # TODO: what's optimal on this?
+    self.assertLess(ops_ratio, 4.3)
+    self.assertLess(mem_ratio, 3)
+
+  def test_dtype(self):
+    IC, OC, X, Y = 4,4,9,9
+    x,w = Tensor.empty(1,IC,Y,X), Tensor.empty(OC,IC,3,3)
+    self.assertEqual(Tensor.conv2d(x,w).dtype, dtypes.default_float)
+
+    x,w = Tensor.empty(1,IC,Y,X,dtype=dtypes.half), Tensor.empty(OC,IC,3,3,dtype=dtypes.half)
+    self.assertEqual(Tensor.conv2d(x,w).dtype, dtypes.half)
+
+if __name__ == '__main__':
+  unittest.main(verbosity=2)

test/opt/test_tensor_cores.py

@@ -149,10 +149,10 @@ class TestTensorCores(unittest.TestCase):
         if _to_np_dtype(real_bufs[0].dtype) is None: continue
         real_bufs[0].copyin(np.zeros((real_bufs[0].size, ), dtype=_to_np_dtype(real_bufs[0].dtype)).data) # Zero to check that all values are filled
         prg.exec(real_bufs)
-        result = np.frombuffer(real_bufs[0].as_buffer(), _to_np_dtype(real_bufs[0].dtype))
+        result = np.frombuffer(real_bufs[0].as_memoryview(), _to_np_dtype(real_bufs[0].dtype))
 
         # ensure the results for each choice of axis matches
-        if golden_result is None: golden_result = np.frombuffer(real_bufs[0].as_buffer(), _to_np_dtype(real_bufs[0].dtype))
+        if golden_result is None: golden_result = np.frombuffer(real_bufs[0].as_memoryview(), _to_np_dtype(real_bufs[0].dtype))
         np.testing.assert_allclose(result, golden_result, atol=0.1, rtol=0.2)
 
   @Context(ALLOW_TF32=1)

test/test_arange.py

@@ -1,7 +1,7 @@
 import unittest
 import numpy as np
 from tinygrad import Tensor, GlobalCounters, dtypes, nn, Device, Variable
-from tinygrad.helpers import Context, getenv
+from tinygrad.helpers import Context, getenv, EMULATE
 from tinygrad.engine.realize import run_schedule
 from tinygrad.engine.realize import CompiledRunner, get_program
 from tinygrad.engine.schedule import ExecItem
@@ -188,5 +188,33 @@ class TestIndexing(unittest.TestCase):
     for i in idx.flatten().numpy(): expected_grad[i] += 2
     np.testing.assert_allclose(emb.weight.grad.numpy(), expected_grad, rtol=1e-5, atol=1e-5)
 
+  # ~10x overhead in fused matmul bw with rope in bf16 vs float16
+  @unittest.skipUnless(Device.DEFAULT == "AMD" or (Device.DEFAULT == "NULL" and EMULATE.value.startswith("AMD")), "tests AMD bf16 cast overhead")
+  def base_test_llama_8b_rope_backward(self, dtype, ops_scale):
+    from extra.models.llama import precompute_freqs_cis, apply_rotary_emb
+    Tensor.training = True
+    bs, seqlen, dim, n_heads = 1, 512, 256, 4
+    head_dim = dim // n_heads
+    x = Tensor.randn(bs, seqlen, dim, dtype=dtype)
+    wq = Tensor.randn(dim, dim, dtype=dtype, requires_grad=True)
+    freqs_cis = precompute_freqs_cis(head_dim, seqlen).cast(dtype)
+    Tensor.realize(x, wq, freqs_cis)
+    xq = (x @ wq.T)
+    # main llama does not fuse it
+    #xq = xq.contiguous_backward()
+    xq = xq.reshape(bs, seqlen, n_heads, head_dim)
+    xq_rope, _ = apply_rotary_emb(xq, xq, freqs_cis)
+    xq_rope.sum().backward()
+    sched = wq.grad.schedule()
+    assert len(sched) == 1, f"expected one kernel for backward, got: {len(sched)}"
+    prg = sched[0].lower().prg.p
+    bwd_ops = prg.estimates.ops
+    expected_ops = bs*seqlen*dim*dim*ops_scale
+    print(f"rope matmul bwd ({dtype}): {GlobalCounters.kernel_count} kernels, {bwd_ops:,} ops")
+    self.assertLess(bwd_ops, expected_ops, f"rope bwd ops {bwd_ops:,} should be < {ops_scale} per (got {bwd_ops/(bs*seqlen*dim*dim):.1f})")
+
+  def test_llama_8b_rope_backward_f16(self): self.base_test_llama_8b_rope_backward(dtypes.float16, 1)
+  def test_llama_8b_rope_backward_bf16(self): self.base_test_llama_8b_rope_backward(dtypes.bfloat16, 11)
+
 if __name__ == "__main__":
   unittest.main()

test/test_const_folding.py

@@ -1,8 +1,7 @@
-import unittest, itertools, math
-from tinygrad import Tensor, Device, dtypes, Context
-from tinygrad.dtype import DType, ConstType, DTYPES_DICT
-from tinygrad.uop.ops import Ops, UOp
-from tinygrad.codegen import full_rewrite_to_sink
+import unittest, math
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.dtype import DTYPES_DICT
+from tinygrad.uop.ops import Ops
 from tinygrad.device import is_dtype_supported
 import numpy as np
 from test.helpers import not_support_multi_device
@@ -13,140 +12,6 @@ def _check_ast_count(desired_count:int, t:Tensor):
   asts = [s for s in schedule if s.ast.op is Ops.SINK]
   assert len(asts) == desired_count, f"{len(asts)} != {desired_count}"
 
-class TestUnaryOpsConstFolding(unittest.TestCase):
-  def test_all_consts_ops(self):
-    _check_ast_count(0, Tensor.ones(4).exp())
-    _check_ast_count(0, Tensor.ones(4).sqrt())
-    _check_ast_count(0, Tensor.ones(4) + Tensor.ones(4))
-    _check_ast_count(0, Tensor.ones(4) / Tensor.ones(4))
-
-  def test_cast(self):
-    _check_ast_count(0, Tensor.ones(4).cast(dtypes.int16))
-    _check_ast_count(0, Tensor.full(4, fill_value=-1).cast(dtypes.uint16))
-
-  def test_neg_folding(self):
-    _check_ast_count(0, Tensor([1, 2, 3]).mul(-1).neg())
-    _check_ast_count(0, Tensor([1, 2, 3]).neg().mul(-1))
-    _check_ast_count(0, Tensor([1, 2, 3]).neg().neg())
-
-  def test_neg_realized_no_fold(self):
-    x = Tensor.randn(32, 32)
-    x = x.clip(0, 1).realize()
-    _check_ast_count(1, x.neg())
-
-class TestBinaryOpsConstFolding(unittest.TestCase):
-  def test_add_literal_zero(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) + 0)
-  def test_add_tensor_zero(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) + Tensor.zeros(4))
-  def test_literal_zero_add(self):
-    _check_ast_count(0, 0 + Tensor([1.0, 2, 3, 4]))
-  def test_tensor_zero_add(self):
-    _check_ast_count(0, Tensor.zeros(4) + Tensor([1.0, 2, 3, 4]))
-
-  def test_sub_literal_zero(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) - 0)
-  def test_sub_tensor_zero(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) - Tensor.zeros(4))
-
-  def test_mul_literal_zero(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) * 0)
-  def test_mul_tensor_zero(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) * Tensor.zeros(4))
-  def test_literal_zero_mul(self):
-    _check_ast_count(0, 0 * Tensor([1.0, 2, 3, 4]) * 0)
-  def test_tensor_zero_mul(self):
-    _check_ast_count(0, Tensor.zeros(4) * Tensor([1.0, 2, 3, 4]))
-
-  def test_mul_literal_one(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) * 1)
-  def test_mul_tensor_one(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) * Tensor.ones(4))
-  def test_literal_one_mul(self):
-    _check_ast_count(0, 1 * Tensor([1.0, 2, 3, 4]))
-  def test_tensor_one_mul(self):
-    _check_ast_count(0, Tensor.ones(4) * Tensor([1.0, 2, 3, 4]))
-
-  def test_bool_tensor_mul_bool(self):
-    _check_ast_count(0, Tensor([True, False]) * True)
-    _check_ast_count(0, Tensor([True, False]) * False)
-  def test_bool_mul_bool_tensor(self):
-    _check_ast_count(0, True * Tensor([True, False]))
-    _check_ast_count(0, False * Tensor([True, False]))
-
-  def test_div_literal_one(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) / 1)
-  def test_div_tensor_one(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) / Tensor.ones(4))
-
-  def test_idiv_literal_one(self):
-    _check_ast_count(0, Tensor([1, 2, 3, 4]) // 1)
-  def test_idiv_tensor_one(self):
-    _check_ast_count(0, Tensor([1, 2, 3, 4]) // Tensor.ones(4, dtype=dtypes.int32))
-
-  def test_pow_literal_zero(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) ** 0)
-  def test_pow_tensor_zero(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) ** Tensor.zeros(4))
-
-  def test_pow_literal_one(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) ** 1)
-  def test_pow_tensor_one(self):
-    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) ** Tensor.ones(4))
-  def test_literal_one_pow(self):
-    _check_ast_count(0, 1 ** Tensor([1.0, 2, 3, 4]))
-  def test_tensor_one_pow(self):
-    _check_ast_count(0, Tensor.ones(4) ** Tensor([1.0, 2, 3, 4]))
-
-class TestBitcastConstFolding(unittest.TestCase):
-  def test_scalar_bitcast(self):
-    def t(cases: dict[DType, ConstType]):
-      for (from_dt, from_v), (to_dt, to_v) in itertools.product(cases.items(), cases.items()):
-        if not math.isnan(from_v):
-          r = full_rewrite_to_sink(UOp.const(from_dt, from_v).bitcast(to_dt).sink()).src[0]
-          self.assertEqual(r.op, Ops.CONST, msg:=f"{from_dt} -> {to_dt} ({from_v} -> {to_v})")
-          self.assertEqual(r.dtype, to_dt, msg)
-          np.testing.assert_equal(r.arg, to_v, msg)
-
-    t({dtypes.int8: 0, dtypes.uint8: 0, dtypes.bool: False})
-    t({dtypes.int8: 1, dtypes.uint8: 1, dtypes.bool: True})
-
-    t({dtypes.int8:  -1, dtypes.uint8:  2**8-1})
-    t({dtypes.int16: -1, dtypes.uint16: 2**16-1, dtypes.float16: float('nan')})
-    t({dtypes.int32: -1, dtypes.uint32: 2**32-1, dtypes.float32: float('nan')})
-    t({dtypes.int64: -1, dtypes.uint64: 2**64-1, dtypes.float64: float('nan')})
-
-    t({dtypes.int8:  -2**7,  dtypes.uint8:  2**7})
-    t({dtypes.int16: -2**15, dtypes.uint16: 2**15})
-    t({dtypes.int32: -2**31, dtypes.uint32: 2**31})
-    t({dtypes.int64: -2**63, dtypes.uint64: 2**63})
-
-    t({dtypes.int16: 13496, dtypes.uint16: 13496, dtypes.float16: 0.294921875})
-    t({dtypes.int32: 1050081145, dtypes.uint32: 1050081145, dtypes.float32: 0.29485681653022766})
-    t({dtypes.int64: 4598983288165178391, dtypes.uint64: 4598983288165178391, dtypes.float64: 0.29485681936461233})
-
-  def test_vec_bitcast(self):
-    with Context(SPEC=0):
-      r = full_rewrite_to_sink(UOp.const(dtypes.int32.vec(3), (-1, -2**31, 75)).bitcast(dtypes.uint32.vec(3)).sink()).src[0]
-    self.assertEqual(r.op, Ops.VECTORIZE)
-    self.assertEqual(r.dtype, dtypes.uint32.vec(3))
-    self.assertEqual(tuple(x.arg for x in r.src), (2**32-1, 2**31, 75))
-
-# folds advance indexing into basic indexing
-class TestIndexingConstFolding(unittest.TestCase):
-  def test_scalar_index(self):
-    t = Tensor.arange(16).float().reshape(1,1,4,4).realize()
-    _check_ast_count(1, t[:,:,Tensor(1),:])
-    _check_ast_count(1, t[:,:,Tensor(1)+2,:])
-    _check_ast_count(1, t[:,:,Tensor(1),Tensor(0)])
-
-  def test_const_tensor_index(self):
-    # TODO: these can be 0, implement const tensor folded indexing
-    t = Tensor.arange(16).float().reshape(1,1,4,4).realize()
-    _check_ast_count(1, t[:,:,Tensor.ones(2,1,dtype=dtypes.int),:])
-    _check_ast_count(1, t[:,:,Tensor.ones(1,2,dtype=dtypes.int)+2,:])
-    _check_ast_count(1, t[:,:,Tensor.ones(1,1,dtype=dtypes.int),Tensor.zeros(2,1,2,dtype=dtypes.int)])
-
 class TestMovedConstFolding(unittest.TestCase):
   def test_add_shrunk_zero(self):
     _check_ast_count(0, Tensor([1.0, 2, 3, 4]) + Tensor.zeros(6).shrink(((1, 5),)))
@@ -169,9 +34,8 @@ class TestMovedConstFolding(unittest.TestCase):
       _check_ast_count(1, Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16))
       np.testing.assert_equal(Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16).numpy(), [0, 65535, 65535, 65535, 65535, 0])
     # folded
-    if is_dtype_supported(dtypes.int64):
-      _check_ast_count(1, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int64))
-      np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int64).numpy(), [0, 1, 1, 1, 1, 0])
+    _check_ast_count(1, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int64))
+    np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int64).numpy(), [0, 1, 1, 1, 1, 0])
 
 class TestReduceOpsConstFolding(unittest.TestCase):
   def test_const_sum(self):

test/test_custom_kernel.py

@@ -52,7 +52,7 @@ def flip_contract_kernel(dest:UOp, src:UOp):
   j = UOp.range(dest.shape[1], 1, AxisType.UPCAST)
   vec = src[i, j].contract(j)
   store = UOp.group(*[dest[i, k].store(vec.gep(3-k)) for k in range(4)])
-  return store.end(i).sink(arg=KernelInfo(name=f"flip_contract_{dest.size}", opts_to_apply=()))
+  return store.end(i, j).sink(arg=KernelInfo(name=f"flip_contract_{dest.size}", opts_to_apply=()))
 
 def slice_sum_kernel(dest:UOp, src:UOp):
   G = UOp.range(src.shape[0], 0)
@@ -88,13 +88,13 @@ def simple_qkv_kernel(O:UOp, Q:UOp, K:UOp, V:UOp) -> UOp:
 # **** backward callbacks ****
 
 def backward_gemm(gradient:UOp, kernel:UOp) -> tuple[UOp, UOp]:
-  out, a, b = kernel.src
+  out, a, b = kernel.src[1:]
   grad_a = (Tensor(gradient) @ Tensor(b).T).uop
   grad_b = (Tensor(a).T @ Tensor(gradient)).uop
   return (None, grad_a, grad_b)
 
 def backward_gemm_custom(gradient:UOp, kernel:UOp) -> tuple[UOp, UOp]:
-  out, a, b = kernel.src
+  out, a, b = kernel.src[1:]
   grad_a = Tensor.empty_like(Tensor(a)).custom_kernel(Tensor(gradient), Tensor(b).T, fxn=custom_gemm)[0].uop
   grad_b = Tensor.empty_like(Tensor(b)).custom_kernel(Tensor(a).T, Tensor(gradient), fxn=custom_gemm)[0].uop
   return (None, grad_a, grad_b)
@@ -104,7 +104,7 @@ def backward_gemm_custom(gradient:UOp, kernel:UOp) -> tuple[UOp, UOp]:
 class TestCustomKernel(unittest.TestCase):
   def test_empty(self):
     a = Tensor.empty(1)
-    a = Tensor.custom_kernel(a, fxn=lambda _: UOp.sink())[0]
+    a = Tensor.custom_kernel(a, fxn=lambda _: UOp.sink(arg=KernelInfo()))[0]
     a.realize()
 
   def test_simple(self):
@@ -128,6 +128,14 @@ class TestCustomKernel(unittest.TestCase):
     out = c.flatten().tolist()
     assert all(x == 2 for x in out), "all 2"
 
+  def test_sharded_add_one(self):
+    # PYTHON backend explicitly checks for OOB access for wrong multi shape regression
+    devs = ("PYTHON:0", "PYTHON:1")
+    a = Tensor.ones(4, 4).contiguous().shard(devs, axis=0)
+    c = Tensor(Tensor.empty(2, 4, device=devs).uop.multi(0), device=devs)
+    c = Tensor.custom_kernel(c, a, fxn=custom_add_one_kernel)[0]
+    assert (c == 2).all().item()
+
   def test_multioutput(self):
     a = Tensor.full((16, 16), 3.).contiguous()
     b = Tensor.full((16, 16), 3.).contiguous()

test/test_dtype.py

@@ -8,7 +8,6 @@ from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float,
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.nir import NIRRenderer
 from tinygrad import Context, Device, Tensor, dtypes
-from tinygrad.uop import Ops
 from hypothesis import given, settings, strategies as strat
 from test.helpers import rand_for_dtype
 from test.unit.test_dtype_spec import _assert_eq, core_dtypes, dtype_ints, dtype_floats, FP8E4M3_MAX, FP8E5M2_MAX
@@ -247,6 +246,17 @@ class TestBFloat16DTypeCast(unittest.TestCase):
 
 class TestHalfDType(TestDType): DTYPE = dtypes.half
 
+class TestEmulatedHalf(TestHalfDType):
+  @classmethod
+  def setUpClass(cls):
+    cls.stack = contextlib.ExitStack()
+    cls.stack.enter_context(Context(EMULATED_DTYPES="half"))
+    cls.DATA = rand_for_dtype(cls.DTYPE, 10)
+
+  @classmethod
+  def tearDownClass(cls): cls.stack.close()
+
+
 class TestFloatDType(TestDType):
   DTYPE = dtypes.float
 
@@ -339,7 +349,6 @@ class TestUint32DType(TestDType): DTYPE = dtypes.uint32
 
 class TestInt64DType(TestDType): DTYPE = dtypes.int64
 
-@unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
 @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
 class TestEmulatedInt64DType(TestInt64DType):
   @classmethod
@@ -356,7 +365,6 @@ class TestUint64DType(TestDType):
   def test_uint64_load(self):
     assert Tensor(2**64 - 1, dtype=dtypes.uint64).numpy() == 2**64 - 1
 
-@unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
 @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
 class TestEmulatedUInt64DType(TestUint64DType):
   @classmethod
@@ -372,9 +380,40 @@ class TestBoolDType(TestDType): DTYPE = dtypes.bool
 
 class TestBFloat16Type(TestDType): DTYPE = dtypes.bfloat16
 
+class TestEmulatedBFloat16Type(TestBFloat16Type):
+  @classmethod
+  def setUpClass(cls):
+    cls.stack = contextlib.ExitStack()
+    cls.stack.enter_context(Context(EMULATED_DTYPES="bfloat16"))
+    cls.DATA = rand_for_dtype(cls.DTYPE, 10)
+
+  @classmethod
+  def tearDownClass(cls): cls.stack.close()
+
 class TestFp8e4m3(TestDType): DTYPE = dtypes.fp8e4m3
+
+class TestEmulatedFp8e4m3(TestFp8e4m3):
+  @classmethod
+  def setUpClass(cls):
+    cls.stack = contextlib.ExitStack()
+    cls.stack.enter_context(Context(EMULATED_DTYPES="fp8e4m3"))
+    cls.DATA = rand_for_dtype(cls.DTYPE, 10)
+
+  @classmethod
+  def tearDownClass(cls): cls.stack.close()
+
 class TestFp8e5m2(TestDType): DTYPE = dtypes.fp8e5m2
 
+class TestEmulatedFp8e5m2(TestFp8e5m2):
+  @classmethod
+  def setUpClass(cls):
+    cls.stack = contextlib.ExitStack()
+    cls.stack.enter_context(Context(EMULATED_DTYPES="fp8e5m2"))
+    cls.DATA = rand_for_dtype(cls.DTYPE, 10)
+
+  @classmethod
+  def tearDownClass(cls): cls.stack.close()
+
 class TestPtrDType(unittest.TestCase):
   def test_vec_double(self):
     dt1 = dtypes.float.vec(4).ptr().vec(4)

test/test_dtype_alu.py

@@ -1,13 +1,12 @@
 import unittest, operator, math
 from tinygrad import Context, Tensor, dtypes, Device
-from tinygrad.dtype import DType, truncate
-from tinygrad.helpers import CI, getenv
+from tinygrad.dtype import DType, truncate, fp8_to_float
+from tinygrad.helpers import CI, EMULATED_DTYPES, getenv
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.device import is_dtype_supported
 from tinygrad.runtime.ops_python import from_storage_scalar
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.nir import NIRRenderer
-from tinygrad.uop import Ops
 import numpy as np
 import pytest
 from hypothesis import assume, given, strategies as strat, settings
@@ -60,11 +59,15 @@ def universal_test(a, b, dtype, op):
   # lt and max with nan is undefined in tinygrad
   if op[0] in (operator.lt, Tensor.maximum) and (math.isnan(a) or math.isnan(b)): return
   ta, tb = Tensor([a], dtype=dtype), Tensor([b], dtype=dtype)
-  tensor_value = (op[0](ta, tb)).numpy()
-  numpy_value = op[1](ta.numpy(), tb.numpy())
-  if dtype in dtypes.fp8s: numpy_value = truncate[dtype](numpy_value.item())
+  if dtype in dtypes.fp8s and op[0] not in (operator.lt, operator.eq):
+    tensor_value = fp8_to_float((op[0](ta.realize(), tb.realize())).bitcast(dtypes.uint8).item(), dtype)
+    numpy_value = truncate[dtype](op[1](ta.numpy(), tb.numpy()).item())
+  else: tensor_value, numpy_value = (op[0](ta, tb)).numpy(), op[1](ta.numpy(), tb.numpy())
   if dtype in dtypes.floats:
-    atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1)}.get(dtype, (1e-10, 1e-7))
+    if not is_dtype_supported(dtype) or dtype in EMULATED_DTYPES.tolist(dtypes): # denormals are zero
+      fe, fm = dtypes.finfo(dtype)
+      atol, rtol = 2 ** (2 - (1 << (fe - 1))), 2 ** (-fm)
+    else: atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1)}.get(dtype, (1e-10, 1e-7))
     np.testing.assert_allclose(tensor_value, numpy_value, atol=atol, rtol=rtol)
   else: np.testing.assert_equal(tensor_value, numpy_value)
 
@@ -74,13 +77,14 @@ def universal_test_unary(a, dtype, op):
   # TODO: cos does not match for large input
   if op[0] == Tensor.cos and abs(a) > 30: return
   if op[0] == Tensor.log and a <= 0: return
-  out: Tensor = op[0](ta)
-  tensor_value = out.numpy()
-  numpy_value = op[1](ta.numpy())
   if dtype in dtypes.fp8s:
+    # normals are zero
+    if dtype in EMULATED_DTYPES.tolist(dtypes) and abs(ta.numpy().item()) < 0.015625: return
+    tensor_value = fp8_to_float(op[0](ta.realize()).bitcast(dtypes.uint8).item(), dtype)
+    numpy_value = truncate[dtype](v:=op[1](ta.numpy()).item())
     # cuda cast f32 inf to f8 MAX, amd cast it to nan(E4M3)/inf(E5M2)
-    if math.isinf(numpy_value.item()): return
-    numpy_value = truncate[dtype](numpy_value.item())
+    if math.isinf(v): return
+  else: tensor_value, numpy_value = op[0](ta).numpy(), op[1](ta.numpy())
   if dtype in dtypes.floats:
     atol, rtol = { dtypes.float16:(1e-3, 1e-2), dtypes.bfloat16:(1e-3, 2e-2),
       dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2: (1.0, 5e-1)}.get(dtype, (1e-6, 1e-5))
@@ -117,21 +121,40 @@ class TestDTypeALU(unittest.TestCase):
   @given(ht.float16, ht.float16, strat.sampled_from(binary_operations))
   def test_float16(self, a, b, op): universal_test(a, b, dtypes.float16, op)
 
+  @given(ht.float16, ht.float16, strat.sampled_from(binary_operations))
+  @Context(EMULATED_DTYPES="half")
+  def test_emulated_float16(self, a, b, op): universal_test(a, b, dtypes.float16, op)
+
   @unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), f"no bfloat16 on {Device.DEFAULT}")
   @given(ht.bfloat16, ht.bfloat16, strat.sampled_from(binary_operations))
   def test_bfloat16(self, a, b, op):
     universal_test(from_storage_scalar(a, dtypes.bfloat16), from_storage_scalar(a, dtypes.bfloat16), dtypes.bfloat16, op)
 
+  @given(ht.bfloat16, ht.bfloat16, strat.sampled_from(binary_operations))
+  @Context(EMULATED_DTYPES="bfloat16")
+  def test_emulated_bfloat16(self, a, b, op):
+    universal_test(from_storage_scalar(a, dtypes.bfloat16), from_storage_scalar(a, dtypes.bfloat16), dtypes.bfloat16, op)
+
   @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3), f"no fp8e4m3 on {Device.DEFAULT}")
   @given(ht.fp8e4m3, ht.fp8e4m3, strat.sampled_from(binary_operations))
   def test_fp8e4m3(self, a, b, op):
     universal_test(from_storage_scalar(a, dtypes.fp8e4m3), from_storage_scalar(b, dtypes.fp8e4m3), dtypes.fp8e4m3, op)
 
+  @given(ht.fp8e4m3, ht.fp8e4m3, strat.sampled_from(binary_operations))
+  @Context(EMULATED_DTYPES="fp8e4m3")
+  def test_emulated_fp8e4m3(self, a, b, op):
+    universal_test(from_storage_scalar(a, dtypes.fp8e4m3), from_storage_scalar(b, dtypes.fp8e4m3), dtypes.fp8e4m3, op)
+
   @unittest.skipUnless(is_dtype_supported(dtypes.fp8e5m2), f"no fp8e5m2 on {Device.DEFAULT}")
   @given(ht.fp8e5m2, ht.fp8e5m2, strat.sampled_from(binary_operations))
   def test_fp8e5m2(self, a, b, op):
     universal_test(from_storage_scalar(a, dtypes.fp8e5m2), from_storage_scalar(b, dtypes.fp8e5m2), dtypes.fp8e5m2, op)
 
+  @given(ht.fp8e5m2, ht.fp8e5m2, strat.sampled_from(binary_operations))
+  @Context(EMULATED_DTYPES="fp8e5m2")
+  def test_emulated_fp8e5m2(self, a, b, op):
+    universal_test(from_storage_scalar(a, dtypes.fp8e5m2), from_storage_scalar(b, dtypes.fp8e5m2), dtypes.fp8e5m2, op)
+
   @given(ht.float32, strat.sampled_from(unary_operations))
   def test_float32_unary(self, a, op): universal_test_unary(a, dtypes.float32, op)
 
@@ -139,22 +162,42 @@ class TestDTypeALU(unittest.TestCase):
   @given(ht.float16, strat.sampled_from(unary_operations))
   def test_float16_unary(self, a, op): universal_test_unary(a, dtypes.float16, op)
 
+  @given(ht.float16, strat.sampled_from(unary_operations))
+  @Context(EMULATED_DTYPES="half")
+  def test_emulated_float16_unary(self, a, op): universal_test_unary(a, dtypes.float16, op)
+
   @unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), f"no bfloat16 on {Device.DEFAULT}")
   @given(ht.bfloat16, strat.sampled_from(unary_operations))
   def test_bfloat16_unary(self, a, op): universal_test_unary(from_storage_scalar(a, dtypes.bfloat16), dtypes.bfloat16, op)
 
+  @given(ht.bfloat16, strat.sampled_from(unary_operations))
+  @Context(EMULATED_DTYPES="bfloat16")
+  def test_emulated_bfloat16_unary(self, a, op): universal_test_unary(from_storage_scalar(a, dtypes.bfloat16), dtypes.bfloat16, op)
+
   @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3), f"no fp8e4m3 on {Device.DEFAULT}")
   @given(ht.fp8e4m3, strat.sampled_from(unary_operations))
   def test_fp8e4m3_unary(self, a, op):
     if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e4m3) != 0.0)
     universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e4m3), dtypes.fp8e4m3, op)
 
+  @given(ht.fp8e4m3, strat.sampled_from(unary_operations))
+  @Context(EMULATED_DTYPES="fp8e4m3")
+  def test_emulated_fp8e4m3_unary(self, a, op):
+    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e4m3) != 0.0)
+    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e4m3), dtypes.fp8e4m3, op)
+
   @unittest.skipUnless(is_dtype_supported(dtypes.fp8e5m2), f"no fp8e5m2 on {Device.DEFAULT}")
   @given(ht.fp8e5m2, strat.sampled_from(unary_operations))
   def test_fp8e5m2_unary(self, a, op):
     if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e5m2) != 0.0)
     universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e5m2), dtypes.fp8e5m2, op)
 
+  @given(ht.fp8e5m2, strat.sampled_from(unary_operations))
+  @Context(EMULATED_DTYPES="fp8e5m2")
+  def test_emulated_fp8e5m2_unary(self, a, op):
+    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e5m2) != 0.0)
+    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e5m2), dtypes.fp8e5m2, op)
+
   @given(ht.uint8, ht.uint8, strat.sampled_from(integer_binary_operations))
   def test_uint8(self, a, b, op): universal_test(a, b, dtypes.uint8, op)
 
@@ -170,7 +213,6 @@ class TestDTypeALU(unittest.TestCase):
   @given(ht.uint64, ht.uint64, strat.sampled_from(integer_binary_operations))
   def test_uint64(self, a, b, op): universal_test(a, b, dtypes.uint64, op)
 
-  @unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
   @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
   @given(ht.uint64, ht.uint64, strat.sampled_from(integer_binary_operations))
   @Context(EMULATED_DTYPES="long")
@@ -185,11 +227,9 @@ class TestDTypeALU(unittest.TestCase):
   @given(ht.int32, ht.int32, strat.sampled_from(integer_binary_operations))
   def test_int32(self, a, b, op): universal_test(a, b, dtypes.int32, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.int64), f"no int64 on {Device.DEFAULT}")
   @given(ht.int64, ht.int64, strat.sampled_from(integer_binary_operations))
   def test_int64(self, a, b, op): universal_test(a, b, dtypes.int64, op)
 
-  @unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
   @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
   @given(ht.int64, ht.int64, strat.sampled_from(integer_binary_operations))
   @Context(EMULATED_DTYPES="long")
@@ -210,7 +250,6 @@ class TestDTypeALU(unittest.TestCase):
   @given(ht.uint64, strat.sampled_from(integer_unary_operations))
   def test_uint64_unary(self, a, op): universal_test_unary(a, dtypes.uint64, op)
 
-  @unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
   @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
   @given(ht.uint64, strat.sampled_from(integer_unary_operations))
   @Context(EMULATED_DTYPES="long")
@@ -225,11 +264,9 @@ class TestDTypeALU(unittest.TestCase):
   @given(ht.int32, strat.sampled_from(integer_unary_operations))
   def test_int32_unary(self, a, op): universal_test_unary(a, dtypes.int32, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.int64), f"no int64 on {Device.DEFAULT}")
   @given(ht.int64, strat.sampled_from(integer_unary_operations))
   def test_int64_unary(self, a, op): universal_test_unary(a, dtypes.int64, op)
 
-  @unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
   @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
   @given(ht.int64, strat.sampled_from(integer_unary_operations))
   @Context(EMULATED_DTYPES="long")

test/test_edgecases.py

@@ -175,7 +175,6 @@ class TestZeroFolding(unittest.TestCase):
 class TestAssignIssues(unittest.TestCase):
   # these are good failures. i'm not sure we need more, but we need to fix these.
 
-  @unittest.expectedFailure
   def test_assign_permuted_view_constant(self):
     # assigning to a permuted view should modify the underlying tensor
     arr = np.arange(6).reshape(2, 3).astype(np.float32)
@@ -185,7 +184,6 @@ class TestAssignIssues(unittest.TestCase):
     t.permute(1, 0).assign(Tensor([[5.0, 6.0], [7.0, 8.0], [9.0, 10.0]]))
     np.testing.assert_allclose(t.numpy(), torch_tensor.numpy())
 
-  @unittest.expectedFailure
   def test_assign_shrink_view_constant(self):
     # assigning to a shrunk view should update the base tensor
     arr = np.arange(9).reshape(3, 3).astype(np.float32)

test/test_fusion_op.py

@@ -1,68 +0,0 @@
-import unittest
-import time
-import numpy as np
-from tinygrad import Tensor, dtypes
-from tinygrad.engine.realize import run_schedule
-
-class TestFusionOp(unittest.TestCase):
-  def test_contiguous_add(self):
-    def test(contig=False):
-      bt = Tensor(np.arange(16), dtype=dtypes.float32).reshape(4,4)
-      x = bt.permute(1,0)
-      if contig: x = x.contiguous()
-      return (x.permute(1,0) + bt).data()
-    assert test() == test(True)
-
-  def test_expand_fuse(self):
-    bt = Tensor(np.ones((10, 1)), dtype=dtypes.float32)
-    out = (bt*2).expand(10,10).sum(1)
-    sched = out.schedule()
-    run_schedule(sched)
-    outd = out.tolist()
-    assert all(x == 20.0 for x in outd)
-
-  def test_recursive_add(self):
-    st = time.perf_counter()
-    a = Tensor([1,2,3,4])
-    for _ in range(24): a = a + a
-    sched = a.schedule()
-    sched[-1].lower()
-    self.assertLess(time.perf_counter()-st, 2.0)
-    assert len(sched[-1].prg.p.src.splitlines()) < 250
-
-  def test_recursive_add_cmp(self):
-    st = time.perf_counter()
-    a = Tensor([1,2,3,4])
-    for _ in range(24): a = a + a
-    sched1 = a.schedule()
-    b = Tensor([1,2,3,4])
-    for _ in range(24): b = b + b
-    sched2 = b.schedule()
-    c = Tensor([1,2,3,4])
-    for _ in range(23): c = c + c
-    sched3 = c.schedule()
-    self.assertEqual(sched1[-1].ast, sched2[-1].ast)
-    with self.assertRaises(AssertionError): self.assertEqual(sched1[-1].ast, sched3[-1].ast)
-    self.assertLess(time.perf_counter()-st, 2.0)
-
-  def test_recursive_pad(self):
-    st = time.perf_counter()
-    val = 1.0
-    a = Tensor(val)
-    for _ in range(24): a = Tensor.stack(a, a)[0]
-    sched = a.schedule()
-    self.assertEqual(len(sched), 0)
-    self.assertLess(time.perf_counter()-st, 2.0)
-
-  def test_recursive_reshape(self):
-    st = time.perf_counter()
-    a = Tensor.empty(32, 32).realize()
-    b = Tensor.empty(16, 2).realize()
-    r = a.sum(1)
-    for _ in range(24): r = r.reshape(16, 2) + b
-    sched = r.schedule()
-    self.assertEqual(len(sched), 1)
-    self.assertLess(time.perf_counter()-st, 2.0)
-
-if __name__ == '__main__':
-  unittest.main(verbosity=2)

test/test_graph.py

@@ -40,7 +40,7 @@ def helper_alloc_rawbuffer(device, fill=False):
   if fill:
     with Context(DEBUG=0):
       data = np.random.randint(-10000, 10000, size=rawbuf.size, dtype=_to_np_dtype(rawbuf.dtype))
-      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_buffer())
+      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_memoryview())
   return rawbuf
 
 def helper_create_offset_rawbuffer(base, offset=0):
@@ -54,7 +54,7 @@ def helper_run_jit(jis, bufs, out_buffers):
     rawbuf.copyin(mv)
 
   for ei in jis: ei.run({}, jit=True)
-  return [rawbuf.as_buffer() for rawbuf in bufs]
+  return [rawbuf.as_memoryview() for rawbuf in bufs]
 
 def helper_test_graphs(graph_impl, graphs, runs=RUN_CNT):
   reg_ji = []

test/test_image_dtype.py

@@ -12,7 +12,7 @@ class TestImageCopy(unittest.TestCase):
   def test_image_copyout_1x8(self, img_type=dtypes.imagef):
     it = Tensor.arange(32).cast(img_type((1,8,4))).realize()
     buf = it.uop.buffer
-    out = buf.as_buffer()
+    out = buf.as_memoryview()
     np.testing.assert_equal(out.cast(it.dtype.fmt).tolist(), np.arange(32))
 
   @unittest.skipUnless(is_dtype_supported(dtypes.half, device="PYTHON"), "need half")
@@ -26,14 +26,14 @@ class TestImageCopy(unittest.TestCase):
   def test_image_copyout_2x4(self):
     it = Tensor.arange(2*4*4).cast(dtypes.imagef((2,4,4))).realize()
     buf = it.uop.buffer
-    out = buf.as_buffer()
+    out = buf.as_memoryview()
     np.testing.assert_equal(out.cast('f').tolist(), np.arange(2*4*4))
 
   def test_image_roundtrip(self):
     sz = (4,2,4)
     it = Tensor.rand(prod(sz)).cast(dtypes.imagef(sz)).realize()
     buf = it.uop.buffer
-    out = buf.as_buffer()
+    out = buf.as_memoryview()
 
     it2 = Tensor.rand(prod(sz)).cast(dtypes.imagef(sz)).realize()
     buf2 = it2.uop.buffer
@@ -190,7 +190,7 @@ class TestImageDType(unittest.TestCase):
       for s in sched:
         s.run()
         if s.bufs[0].dtype == dtypes.float:
-          lst = s.bufs[0].as_buffer().cast("f").tolist()
+          lst = s.bufs[0].as_memoryview().cast("f").tolist()
           print(lst)
           assert not np.any(np.isnan(lst))
       # NOTE: the w1 grad must realize to a separate kernel

test/test_jit_footguns.py

@@ -111,30 +111,18 @@ class TestJitFootguns(unittest.TestCase):
       self.assertEqual(first.numpy().item(), expected_first)
       buf = new_buf
 
-  def test_slice_assign_requires_realize(self):
-    """Slice assign then read from same buffer - assign isn't connected to read without explicit realize()."""
+  def test_slice_assign_works_without_realize(self):
+    """Slice assign then read from same buffer - pending assigns are side-realized."""
     from tinygrad import Variable
     v_pos = Variable("pos", 0, 3)
-
-    # without .realize() after assign, the read doesn't see the assigned values
     cache = Tensor.zeros(4, 4).contiguous().realize()
     @TinyJit
-    def f_broken(pos):
+    def f(pos):
       cache[pos:pos+1, :].assign(Tensor.ones(1, 4))
       return cache.sum().realize()
     for i in range(4):
       cache.assign(Tensor.zeros(4, 4)).realize()
-      self.assertEqual(f_broken(v_pos.bind(i)).item(), 0.0)  # should be 4.0!
-
-    # workaround: add .realize() after assign
-    cache2 = Tensor.zeros(4, 4).contiguous().realize()
-    @TinyJit
-    def f_fixed(pos):
-      cache2[pos:pos+1, :].assign(Tensor.ones(1, 4)).realize()
-      return cache2.sum().realize()
-    for i in range(4):
-      cache2.assign(Tensor.zeros(4, 4)).realize()
-      self.assertEqual(f_fixed(v_pos.bind(i)).item(), 4.0)
+      self.assertEqual(f(v_pos.bind(i)).item(), 4.0)
 
   def test_symbolic_pad_view_frozen(self):
     """Symbolic pad view has BIND values baked in at capture time. TODO: pad should be captured in jit."""

test/test_linearizer.py

@@ -504,7 +504,7 @@ def helper_linearizer_opt(r:Tensor|list[Tensor], *args, **kwargs):
   return realized_ast
 
 def copyout_outputs(outbufs:list[Buffer]) -> list[np.ndarray]:
-  return [np.frombuffer(x.as_buffer(), _to_np_dtype(x.dtype)) for x in outbufs]
+  return [np.frombuffer(x.as_memoryview(), _to_np_dtype(x.dtype)) for x in outbufs]
 
 def reset_bufs(bufs:list[Buffer]):
   for buf in bufs: buf.copyin(np.zeros((buf.size*buf.dtype.itemsize,), dtype=np.uint8).data)

test/test_multitensor.py

@@ -1248,58 +1248,6 @@ class TestTensorOps(unittest.TestCase):
   def test_bitcast(self):
     helper_test_shard_op([(256,), (256,)], lambda x: x.bitcast(dtypes.int))
 
-@unittest.skipIf(not_support_multi_device(), "no multi")
-class TestMultiRamUsage(unittest.TestCase):
-  def setUp(self):
-    self.baseline = GlobalCounters.mem_used
-    self.N = 100
-  def assertUsed(self, amt, strict=True):
-    used = GlobalCounters.mem_used - self.baseline
-    print(f"used {used} bytes")
-    if strict: self.assertEqual(used, amt)
-    else: self.assertLessEqual(used, amt)
-
-  def test_zeros(self):
-    _ = Tensor.zeros(self.N, self.N).contiguous().realize()
-    self.assertUsed(self.N*self.N*4)
-
-  def test_zeros_del(self):
-    _ = Tensor.zeros(self.N, self.N).contiguous().realize()
-    del _
-    self.assertUsed(0)
-
-  @unittest.skip("flaky")
-  def test_zeros_copy(self):
-    _ = Tensor.zeros(self.N, self.N).contiguous().to(devices_2).realize()
-    # NOTE: the first one on the DEFAULT device should be freed
-    self.assertUsed(self.N*self.N*4*2)
-
-  @unittest.skip("flaky")
-  def test_zeros_shard(self, devices=(d1, d2)):
-    _ = Tensor.zeros(self.N, self.N).contiguous().shard(devices, axis=0).realize()
-    self.assertUsed(self.N*self.N*4) # sharding should not increase total ram usage
-  def test_zeros_shard_self(self): self.test_zeros_shard((d0, d1))
-
-  @unittest.skip("flaky")
-  def test_zeros_contiguous_shard(self):
-    _ = Tensor.zeros(self.N, self.N).contiguous().shard(devices_2, axis=0).contiguous().realize()
-    self.assertUsed(self.N*self.N*4) # sharding should not increase total ram usage
-
-  def _test_matmul_half(self, dev_count:int):
-    N = 32
-    total_mem = {}
-    devs = tuple(f"NULL:{i}" for i in range(dev_count))
-    for dtype in {dtypes.float, dtypes.half}:
-      GlobalCounters.reset()
-      a = Tensor.empty((N, N), dtype=dtype, device=devs[0]).shard(devs, axis=0)
-      b = Tensor.empty((N, N), dtype=dtype, device=devs[0]).shard(devs, axis=None)
-      (a @ b).realize()
-      total_mem[dtype] = GlobalCounters.global_mem
-    self.assertEqual(total_mem[dtypes.half], total_mem[dtypes.float] // 2)
-
-  def test_matmul_half(self): self._test_matmul_half(dev_count=2)
-  def test_matmul_half_alt(self): self._test_matmul_half(dev_count=4)
-
 @unittest.skipIf(not_support_multi_device(), "need multi")
 class TestMultiFromUnrenderable(unittest.TestCase):
   @needs_second_gpu

test/test_ops.py

@@ -596,7 +596,7 @@ class TestOps(unittest.TestCase):
     helper_test_op(None, lambda x: x//2, forward_only=True, vals=[[3, 4, 5]])
     helper_test_op(None, functools.partial(torch.div, rounding_mode="trunc"), Tensor.idiv, forward_only=True,
                    vals=[[-4, 7, 5, 4, -7, 8], [2, -3, 8, -2, 3, 5]])
-    if is_dtype_supported(dtypes.uint64) and not COMPILE_ONLY:
+    if not COMPILE_ONLY:
       x = Tensor(2**64 - 1, dtype=dtypes.uint64).idiv(1)
       np.testing.assert_equal(x.numpy(), 2**64 - 1)
 

test/test_outerworld.py

@@ -38,7 +38,7 @@ class TestOuterRange(unittest.TestCase):
     vi = UOp.variable("i", i.vmin, i.vmax).bind(i)
     out = Tensor(acc.uop.after(acc_i.store(acc_i + a[:, vi].uop).end(i)))
     out.realize()
-    assert all(x == 10.0 for x in out.tolist())
+    self.assertEqual(out.tolist(), [10.0]*10)
 
   def test_range_matmul(self):
     vec = Tensor.randn(1, 10).realize()
@@ -57,7 +57,7 @@ class TestOuterRange(unittest.TestCase):
     out.realize()
 
     # TODO: testing allclose
-    assert Tensor.allclose(ref, out, atol=1e-6), f"{ref.numpy()=}, {out.numpy()=}"
+    assert Tensor.allclose(ref, out, atol=1e-6), f"max diff {(ref-out).abs().max().item()}"
 
 class TestOuterScan(unittest.TestCase):
   def _test_scan(self):
@@ -85,7 +85,7 @@ class TestOuterScan(unittest.TestCase):
     out.realize()
 
     # TODO: testing allclose
-    assert Tensor.allclose(ref, out, atol=1e-6), f"{ref.numpy()=}, {out.numpy()=}"
+    assert Tensor.allclose(ref, out, atol=1e-5), f"max diff {(ref-out).abs().max().item()}"
 
 class TestOuterworld(unittest.TestCase):
   def test_range_plus_1(self):
@@ -166,7 +166,7 @@ class TestVmap(unittest.TestCase):
     out.realize()
 
     # TODO: testing allclose
-    assert Tensor.allclose(ref, out, atol=1e-6), f"{ref.numpy()=}, {out.numpy()=}"
+    assert Tensor.allclose(ref, out, atol=1e-6), f"max diff {(ref-out).abs().max().item()}"
   def test_vmap_inner_fuse(self): self.test_vmap_inner(fuse=True)
   def test_vmap_outer(self): self.test_vmap_inner(AxisType.OUTER)
   def test_vmap_outer_fuse(self): self.test_vmap_inner(AxisType.OUTER, fuse=True)

test/test_pickle.py

@@ -76,7 +76,7 @@ class TestPickle(unittest.TestCase):
     del a
     del buffer
     a2:UOp = pickle.loads(s)
-    self.assertListEqual(a2.base.realized.as_buffer().cast("I").tolist(), [0, 1, 2, 3])
+    self.assertListEqual(a2.base.realized.as_memoryview().cast("I").tolist(), [0, 1, 2, 3])
 
   def test_pickle_unrealized_tensor(self):
     t = Tensor.ones(10, 10)

test/test_profiler.py

@@ -6,6 +6,9 @@ from tinygrad.runtime.support.hcq import HCQCompiled
 from tinygrad.engine.realize import get_runner
 
 MOCKGPU = getenv("MOCKGPU")
+def _dev_base(d):
+  p = d.split(":")
+  return p[0] if len(p) < 2 or not p[1].isdigit() else f"{p[0]}:{p[1]}"
 
 @contextlib.contextmanager
 def helper_collect_profile(*devs):
@@ -55,7 +58,7 @@ class TestProfiler(unittest.TestCase):
     kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent)]
     assert len(kernel_runs) == 1, "one kernel run is expected"
     assert kernel_runs[0].name == runner_name, "kernel name is not correct"
-    assert not kernel_runs[0].is_copy, "kernel should not be copy"
+    assert _dev_base(kernel_runs[0].device) == kernel_runs[0].device, "kernel should not be on a sub-device"
 
   def test_profile_copyin(self):
     buf1 = Buffer(Device.DEFAULT, 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
@@ -63,10 +66,8 @@ class TestProfiler(unittest.TestCase):
     with helper_collect_profile(TestProfiler.d0) as profile:
       buf1.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
 
-    profile, _ = helper_profile_filter_device(profile, TestProfiler.d0.device)
-    kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent)]
+    kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith(TestProfiler.d0.device)]
     assert len(kernel_runs) == 1, "one kernel run is expected"
-    assert kernel_runs[0].is_copy, "kernel should be copy"
 
   def test_profile_multiops(self):
     runner_name = TestProfiler.runner._prg.name
@@ -77,16 +78,12 @@ class TestProfiler(unittest.TestCase):
       TestProfiler.runner([buf1, TestProfiler.a.uop.buffer], var_vals={})
       buf1.copyout(memoryview(bytearray(buf1.nbytes)))
 
-    profile, _ = helper_profile_filter_device(profile, TestProfiler.d0.device)
-    evs = [x for x in profile if isinstance(x, ProfileRangeEvent)]
+    evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith(TestProfiler.d0.device)]
 
     assert len(evs) == 3, "3 kernel runs are expected"
     # NOTE: order of events does not matter, the tool is responsible for sorting them
-    copy_events = [e for e in evs if e.is_copy]
-    self.assertEqual(len(copy_events), 2)
-
-    prg_events = [e for e in evs if not e.is_copy]
-    assert prg_events[0].name == runner_name, "kernel name is not correct"
+    prg_events = [e for e in evs if e.device == TestProfiler.d0.device]
+    assert any(e.name == runner_name for e in prg_events), "kernel name is not correct"
 
     #for i in range(1, 3):
     #  assert evs[i].st > evs[i-1].en, "timestamp not aranged"
@@ -102,13 +99,9 @@ class TestProfiler(unittest.TestCase):
       buf1.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
       buf2.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
 
-    profile0, _ = helper_profile_filter_device(profile, TestProfiler.d0.device)
-    profile1, _ = helper_profile_filter_device(profile, d1.device)
-
-    for p in [profile0, profile1]:
-      evs = [x for x in p if isinstance(x, ProfileRangeEvent)]
+    for dev in [TestProfiler.d0.device, d1.device]:
+      evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and _dev_base(x.device) == dev]
       assert len(evs) == 1, "one kernel runs are expected"
-      assert evs[0].is_copy, "kernel should be copy"
 
   def test_profile_multidev_transfer(self):
     try: d1 = Device[f"{Device.DEFAULT}:1"]
@@ -118,10 +111,8 @@ class TestProfiler(unittest.TestCase):
     with helper_collect_profile(TestProfiler.d0, d1) as profile:
       buf1.to(f"{Device.DEFAULT}:1").realize()
 
-    profile0, _ = helper_profile_filter_device(profile, TestProfiler.d0.device)
-    kernel_runs = [x for x in profile0 if isinstance(x, ProfileRangeEvent)]
+    kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith(TestProfiler.d0.device)]
     assert len(kernel_runs) == 1, "one kernel run is expected"
-    assert kernel_runs[0].is_copy, "kernel should be copy"
 
   @unittest.skipIf(Device.DEFAULT in "METAL" or (MOCKGPU and Device.DEFAULT == "AMD"), "AMD mockgpu does not support queue wait interrupts")
   def test_profile_graph(self):
@@ -167,17 +158,19 @@ class TestProfiler(unittest.TestCase):
       return d2.timeline_signal.timestamp - d1.timeline_signal.timestamp
 
     # then test it by timing the GPU to GPU times
+    dev_evs = {x.device:x for x in Compiled.profile_events if isinstance(x, ProfileDeviceEvent)}
     jitter_matrix = [[float('nan')] * len(devs) for _ in range(len(devs))]
     pairs = [(p1, p2) for p1 in enumerate(devs) for p2 in enumerate(devs) if p1 != p2]
     for (i1, d1), (i2, d2) in pairs:
-      cpu_diff = d1.gpu2cpu_compute_time_diff - d2.gpu2cpu_compute_time_diff
+      cpu_diff = dev_evs[d1.device].tdiff - dev_evs[d2.device].tdiff
       jitter_matrix[i1][i2] = statistics.median(_sync_d2d(d1, d2) - _sync_d2d(d2, d1) for _ in range(20)) / 2 - cpu_diff
 
+    print("pairwise clock jitter matrix (us):\n" + '\n'.join([''.join([f'{float(item):8.3f}' for item in row]) for row in jitter_matrix]))
+
     for (i1, d1), (i2, d2) in pairs:
       assert abs(jitter_matrix[i1][i2]) < 0.5, "jitter should be less than 0.5us"
 
-    print("pairwise clock jitter matrix (us):\n" + '\n'.join([''.join([f'{float(item):8.3f}' for item in row]) for row in jitter_matrix]))
-
+  @unittest.skip("this test is flaky")
   def test_cpu_profile(self):
     def test_fxn(err=False):
       time.sleep(0.1)
@@ -219,6 +212,7 @@ class TestProfiler(unittest.TestCase):
     for ge in graphs:
       self.assertEqual(len(ge.ents), len(graphs))
 
+  @unittest.skip("this test is flaky")
   def test_trace_metadata(self):
     with Context(TRACEMETA=1):
       a = Tensor.empty(1)+2
@@ -227,7 +221,7 @@ class TestProfiler(unittest.TestCase):
         Tensor.realize(a, b)
     profile, _ = helper_profile_filter_device(profile, TestProfiler.d0.device)
     exec_points = [e for e in profile if isinstance(e, ProfilePointEvent) and e.name == "exec"]
-    range_events = [e for e in profile if isinstance(e, ProfileRangeEvent) and not e.is_copy]
+    range_events = [e for e in profile if isinstance(e, ProfileRangeEvent) and _dev_base(e.device) == e.device]
     self.assertEqual(len(exec_points), len(range_events), 2)
     self.assertEqual(len(dedup(e.arg['name'] for e in exec_points)), 1)
     self.assertEqual(len(dedup(e.arg['metadata'] for e in exec_points)), 1)

test/test_randomness.py

@@ -83,7 +83,7 @@ class TestRandomness(unittest.TestCase):
     self.assertTrue(r1.uop.is_realized, "tensor should be realized after .realize()")
     self.assertTrue(r2.uop.is_realized, "tensor should be realized after .realize()")
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.float16) and is_dtype_supported(dtypes.ulong), "need float16 and ulong support")
+  @unittest.skipUnless(is_dtype_supported(dtypes.float16), "need float16 support")
   def test_rand_float16(self):
     N = 128
     x = Tensor.rand((2, N, N), dtype=dtypes.float16)