Merge branch 'master' into precompile_backward

combine no_vectorized_index and no_vectorized_index_broadcast

.github/actions/setup-tinygrad/action.yml

@@ -45,6 +45,10 @@ inputs:
     description: "Install mesa"
     required: false
     default: 'false'
+  tinydreno:
+    description: "Install tinydreno"
+    required: false
+    default: 'false'
 runs:
   using: "composite"
   steps:
@@ -62,14 +66,14 @@ runs:
       uses: actions/cache/restore@v4
       with:
         path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
     - name: Cache Python packages
       if: github.event_name != 'pull_request'
       id: restore-venv
       uses: actions/cache@v4
       with:
         path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
 
     # **** Caching downloads ****
 
@@ -195,13 +199,13 @@ runs:
       uses: actions/cache/restore@v4
       with:
         path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
     - name: Cache apt
       if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name != 'pull_request'
       uses: actions/cache@v4
       with:
         path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
 
     - name: Run apt Update + Install
       if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
@@ -326,3 +330,9 @@ runs:
       if: inputs.mesa == 'true' && runner.os == 'macOS'
       shell: bash
       run: brew install sirhcm/tinymesa/tinymesa_cpu
+
+    # *** tinydreno ***
+    - name: Install tinydreno (linux)
+      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so

.github/workflows/test.yml

@@ -1,7 +1,7 @@
 name: Unit Tests
 env:
   # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '17'
+  CACHE_VERSION: '18'
   CAPTURE_PROCESS_REPLAY: 1
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   PYTHONPATH: ${{ github.workspace }}
@@ -1011,3 +1011,26 @@ jobs:
           python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
           DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
           python -m pytest -n=auto test/backend/test_ops.py --durations=20
+  qcomclcompiletests:
+    name: Compile-only (QCOM CL)
+    runs-on: ubuntu-24.04-arm
+    timeout-minutes: 15
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v4
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: compile-qcomcl
+          deps: testing_unit
+          tinydreno: 'true'
+          python-version: '3.12'
+      - name: Set env
+        shell: bash
+        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\nNULL_QCOMCL=1" >> $GITHUB_ENV
+      - name: Run test_ops
+        shell: bash
+        run: |
+          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
+          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
+          python -m pytest -n=auto test/backend/test_ops.py --durations=20

examples/mlperf/model_train.py

@@ -1282,7 +1282,7 @@ def train_bert():
         previous_step = i
 
 def train_llama3():
-  from extra.models.llama import Transformer
+  from examples.mlperf.models.llama import Transformer
   from examples.llama3 import MODEL_PARAMS
   from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
   from examples.mlperf.optim import GradAccClipAdamW
@@ -1343,7 +1343,8 @@ def train_llama3():
   if (MP := getenv("MP", 1)) > 1: model_params['vocab_size'] = round_up(model_params['vocab_size'], 256 * MP)
   vocab_mask:Tensor = Tensor.arange(model_params['vocab_size']).reshape(1, 1, -1) >= real_vocab_size
 
-  model = Transformer(**model_params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
+  model = Transformer(**model_params, max_context=SEQLEN)
+
   params = get_parameters(model)
   # weights are all bfloat16 for now
   assert params and all(p.dtype == dtypes.bfloat16 for p in params)
@@ -1381,16 +1382,20 @@ def train_llama3():
 
     vocab_mask.shard_(device, axis=2).realize()
 
-  optim_device = "CPU" if getenv("OFFLOAD_OPTIM") else None
+  is_offload_optim = bool(getenv("OFFLOAD_OPTIM"))
+  is_fake_offload = Device.DEFAULT == "NULL"
+  optim_device = ("CPU" if not is_fake_offload else "NULL:99") if is_offload_optim else None
   optim = GradAccClipAdamW(get_parameters(model), lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2,
                            eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device)
 
   # init grads
-  for p in optim.params:
-    p.grad = p.empty_like().realize()
+  if is_offload_optim:
+    for p in optim.params:
+      p.grad = Tensor.zeros(p.shape, dtype=p.dtype, device=optim_device, requires_grad=False).contiguous().realize()
+  else:
+    for p in optim.params:
+      p.grad = p.zeros_like().contiguous().realize()
   grads: list[Tensor] = [p.grad for p in optim.params]
-  for p in optim.params:
-    p.grad.assign(p.grad.zeros_like()).realize()
 
   scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)
 
@@ -1412,12 +1417,13 @@ def train_llama3():
       device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
       tokens = tokens.shard(device)
     if DP == 1 and MP == 1: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
+    logits:Tensor = model(tokens[:, :-1])
     loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
     loss.backward()
     assert all(p.grad is g for p,g in zip(optim.params, grads))
-    Tensor.realize(loss, *grads)
-    return loss.flatten().float().to("CPU")
+    loss_cpu = loss.flatten().float().to("CPU")
+    Tensor.realize(loss_cpu, *grads)
+    return loss_cpu
 
   @TinyJit
   def optim_step():
@@ -1425,12 +1431,13 @@ def train_llama3():
     scheduler.step()
 
     for g in grads:
-      g.assign(g.zeros_like()).realize()
+      g.assign(g.zeros_like())
 
-    lr = optim.lr
-    Tensor.realize(lr, *grads)
+    lr_cpu = optim.lr.float().to("CPU")
+    grad_norm_cpu = grad_norm.float().to("CPU")
+    Tensor.realize(lr_cpu, grad_norm_cpu, *grads)
 
-    return lr.float().to("CPU"), grad_norm.float().to("CPU")
+    return lr_cpu, grad_norm_cpu
 
   @TinyJit
   @Tensor.train(False)
@@ -1442,7 +1449,7 @@ def train_llama3():
       device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
       tokens = tokens.shard(device)
     if DP == 1 and MP == 1: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
+    logits:Tensor = model(tokens[:, :-1])
     loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
     return loss.flatten().float().to("CPU")
 

examples/mlperf/models/llama.py

@@ -0,0 +1,80 @@
+from tinygrad import Tensor, nn
+from tinygrad.helpers import getenv
+from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
+
+class Attention:
+  def __init__(self, dim:int, n_heads:int, n_kv_heads:int|None=None, linear=nn.Linear):
+    self.n_heads = n_heads
+    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
+    self.head_dim = dim // n_heads
+    self.n_rep = self.n_heads // self.n_kv_heads
+
+    if getenv("WQKV"):
+      self.wqkv = linear(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2, bias=False)
+    else:
+      self.wq = linear(dim, self.n_heads * self.head_dim, bias=False)
+      self.wk = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
+      self.wv = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
+
+    self.wo = linear(self.n_heads * self.head_dim, dim, bias=False)
+
+  def __call__(self, x:Tensor, freqs_cis:Tensor) -> Tensor:
+    if getenv("WQKV"):
+      xqkv = self.wqkv(x)
+      xqkv = xqkv.reshape(xqkv.shape[0], xqkv.shape[1], self.n_kv_heads, self.n_rep + 2, self.head_dim)
+      xq = xqkv[:, :, :, :self.n_rep].reshape(xqkv.shape[0], xqkv.shape[1], -1)
+      xk = xqkv[:, :, :, self.n_rep:self.n_rep+1].reshape(xqkv.shape[0], xqkv.shape[1], -1)
+      xv = xqkv[:, :, :, self.n_rep+1:self.n_rep+2].reshape(xqkv.shape[0], xqkv.shape[1], -1)
+    else:
+      xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
+
+    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)
+
+    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
+    bsz, seqlen, _, _ = xq.shape
+
+    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
+    attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
+
+    attn = attn.reshape(bsz, seqlen, -1)
+    return self.wo(attn)
+
+class FeedForward:
+  def __init__(self, dim:int, hidden_dim:int, linear=nn.Linear):
+    self.w1 = linear(dim, hidden_dim, bias=False)
+    self.w2 = linear(hidden_dim, dim, bias=False)
+    self.w3 = linear(dim, hidden_dim, bias=False) # the gate in Gated Linear Unit
+
+  def __call__(self, x:Tensor) -> Tensor:
+    w1 = self.w1(x).silu()
+    w3 = self.w3(x)
+    return self.w2(w1 * w3)
+
+class TransformerBlock:
+  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int|None, norm_eps:float, linear=nn.Linear):
+    self.attention = Attention(dim, n_heads, n_kv_heads, linear)
+    self.feed_forward = FeedForward(dim, hidden_dim, linear)
+    self.attention_norm = nn.RMSNorm(dim, norm_eps)
+    self.ffn_norm = nn.RMSNorm(dim, norm_eps)
+
+  def __call__(self, x:Tensor, freqs_cis:Tensor):
+    h = x + self.attention(self.attention_norm(x), freqs_cis)
+    return h + self.feed_forward(self.ffn_norm(h))
+
+class Transformer:
+  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
+               rope_theta:int=10000, max_context:int=1024, linear=nn.Linear, embedding=nn.Embedding):
+    self.layers = [TransformerBlock(dim, hidden_dim, n_heads, n_kv_heads, norm_eps, linear) for _ in range(n_layers)]
+    self.norm = nn.RMSNorm(dim, norm_eps)
+    self.tok_embeddings = embedding(vocab_size, dim)
+    self.output = nn.Linear(dim, vocab_size, bias=False) if embedding == nn.Embedding else linear(dim, vocab_size, bias=False)
+    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
+
+  def __call__(self, tokens:Tensor):
+    h = self.tok_embeddings(tokens)
+    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
+    for layer in self.layers: h = layer(h, freqs_cis)
+    logits = self.output(self.norm(h))
+    return logits

examples/mlperf/optim.py

@@ -34,10 +34,10 @@ class GradAccClipAdamW(Optimizer):
       grads[0].assign((grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype))
     else:
       for i in range(len(grads)):
-        grads[i].assign(grads[i] / self.grad_acc).realize()
-      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous().realize()
+        grads[i].assign(grads[i] / self.grad_acc)
+      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
       for i in range(len(grads)):
-        grads[i].assign((grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype)).realize()
+        grads[i].assign((grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype))
 
     ret = []
     self.b1_t *= self.b1
@@ -45,8 +45,8 @@ class GradAccClipAdamW(Optimizer):
     for i, g in enumerate(grads):
       self.m[i].assign((self.b1 * self.m[i] + (1.0 - self.b1) * g).cast(self.m[i].dtype))
       self.v[i].assign((self.b2 * self.v[i] + (1.0 - self.b2) * (g * g)).cast(self.v[i].dtype))
-      m_hat = self.m[i] / (1.0 - self.b1_t)
-      v_hat = self.v[i] / (1.0 - self.b2_t)
+      m_hat = (self.m[i] / (1.0 - self.b1_t)).cast(self.m[i].dtype)
+      v_hat = (self.v[i] / (1.0 - self.b2_t)).cast(self.v[i].dtype)
       up = m_hat / (v_hat.sqrt() + self.eps)
       ret.append((self.lr * up).cast(g.dtype))
     return ret, [self.b1_t, self.b2_t] + self.m + self.v + [total_norm]

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama405b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -12,6 +12,7 @@ export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-0}
 export ASM_GEMM=${ASM_GEMM:-1}
 export WQKV=${WQKV:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-1} MP=${MP:-8}

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama405b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -12,6 +12,7 @@ export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-0}
 export ASM_GEMM=${ASM_GEMM:-1}
 export WQKV=${WQKV:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-1} MP=${MP:-8}

extra/amdpci/am_smi.py

@@ -65,7 +65,7 @@ def get_bar0_size(pcibus):
 class AMSMI(AMDev):
   def __init__(self, pcibus, vram_bar:MMIOInterface, doorbell_bar:MMIOInterface, mmio_bar:MMIOInterface):
     self.pcibus = pcibus
-    self.vram, self.doorbell64, self.mmio, self.dma_regions = vram_bar, doorbell_bar, mmio_bar, None
+    self.vram, self.doorbell64, self.mmio = vram_bar, doorbell_bar, mmio_bar
     self.pci_state = self.read_pci_state()
     if self.pci_state == "D0": self._init_from_d0()
 

extra/amdpci/hive_reset.py

@@ -2,7 +2,7 @@
 
 import os
 from tinygrad.helpers import Context
-from tinygrad.runtime.support.system import System, PCIDevice, PCIDevImplBase
+from tinygrad.runtime.support.system import System, PCIDevice
 from tinygrad.runtime.support.hcq import FileIOInterface
 from tinygrad.runtime.support.am.amdev import AMDev
 
@@ -12,7 +12,7 @@ if __name__ == "__main__":
     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]
+  pcidevs = [PCIDevice("AM", gpu) for gpu in gpus]
   amdevs = []
   with Context(DEBUG=2):
     for pcidev in pcidevs:

extra/amdpci/proclogs.py

@@ -7,8 +7,8 @@ class GFXFake:
   def __init__(self): self.xccs = 8
 
 class AMDFake(AMDev):
-  def __init__(self, pci_dev, dma_regions=None):
-    self.pci_dev, self.devfmt, self.dma_regions = pci_dev, pci_dev.pcibus, dma_regions
+  def __init__(self, pci_dev):
+    self.pci_dev, self.devfmt = pci_dev, pci_dev.pcibus
     self.vram, self.doorbell64, self.mmio = self.pci_dev.map_bar(0), self.pci_dev.map_bar(2, fmt='Q'), self.pci_dev.map_bar(5, fmt='I')
     self._run_discovery()
     self._build_regs()

extra/gemm/amd_copy_matmul.py

@@ -0,0 +1,84 @@
+from tinygrad import UOp, getenv
+from tinygrad.uop.ops import AxisType, KernelInfo, Ops
+from tinygrad.dtype import AddrSpace, dtypes
+
+N = getenv("N", 4096)
+M = getenv("M", N)
+K = getenv("K", N)
+
+WARP_SIZE = 32
+BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 8
+TM, TN = 4, 4
+LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8
+assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0
+
+is_kernel5 = getenv("K5", 0)
+THREADS_PER_BLOCK = 128 if is_kernel5 else 256
+WAVES_PER_BLOCK_N = 1 if is_kernel5 else 2
+WAVES_PER_BLOCK_M = THREADS_PER_BLOCK // WARP_SIZE // WAVES_PER_BLOCK_N
+REG_TILES_PER_WAVE_N = BLOCK_N // (WAVES_PER_BLOCK_N * LANES_PER_WAVE_N * TN)
+REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM)
+
+assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong"
+assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong"
+
+consts = {"wpb_m":WAVES_PER_BLOCK_M, "lpw_m":LANES_PER_WAVE_M, "rt_m":REG_TILES_PER_WAVE_M, "t_m": TM,
+          "wpb_n":WAVES_PER_BLOCK_N, "lpw_n":LANES_PER_WAVE_N, "rt_n":REG_TILES_PER_WAVE_N, "t_n": TN}
+
+# 128x128 out, kx128, kx128 in
+def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
+  tid = UOp.range(THREADS_PER_BLOCK, 2, AxisType.LOCAL)
+
+  #tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
+  warp, lane = tid // WARP_SIZE, tid % WARP_SIZE
+  wave_n, wave_m = warp % WAVES_PER_BLOCK_N, warp // WAVES_PER_BLOCK_N
+  lane_n, lane_m = lane % LANES_PER_WAVE_N, lane // LANES_PER_WAVE_N
+
+  # define locals
+  A_local = UOp.placeholder((BLOCK_K, BLOCK_M), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL)
+  B_local = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
+
+  # open the main reduction range and copy in GLOBAL -> LOCAL
+  a = a.reshape(K // BLOCK_K, BLOCK_K, BLOCK_M)
+  b = b.reshape(K // BLOCK_K, BLOCK_K, BLOCK_N)
+  k_tile_range = UOp.range(K // BLOCK_K, 3, AxisType.REDUCE)
+  A_store = A_local.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(a[k_tile_range].reshape(-1, THREADS_PER_BLOCK)[:, tid])
+  B_store = B_local.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(b[k_tile_range].reshape(-1, THREADS_PER_BLOCK)[:, tid])
+  barrier = UOp.barrier(A_store, B_store)
+  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
+
+  # define accumulator (128x128), but broadcast across tid
+  c_regs = UOp.placeholder((REG_TILES_PER_WAVE_M*TM, REG_TILES_PER_WAVE_N*TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  c_regs = c_regs.after(c_regs.store(UOp.const(dtypes.float, 0).reshape((1,)*len(c_regs.shape)).expand(c_regs.shape)))
+
+  # define registers (NOTE: the thread count is the device count for this multi, it's sharded across the THREADS_PER_BLOCK)
+  A_col = UOp.placeholder((REG_TILES_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
+  B_row = UOp.placeholder((REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
+
+  # LOCAL -> REGS
+  k = UOp.range(BLOCK_K, 4, AxisType.REDUCE)
+  A_col = A_col.after(A_col.store(A_local[k].reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM)[wave_m, :, lane_m, :]))
+  B_row = B_row.after(B_row.store(B_local[k].reshape(WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)[wave_n, :, lane_n, :]))
+
+  # do FMA
+  A_col = A_col.reshape(REG_TILES_PER_WAVE_M*TM, 1).expand(REG_TILES_PER_WAVE_M*TM, REG_TILES_PER_WAVE_N*TN)
+  B_row = B_row.reshape(1, REG_TILES_PER_WAVE_N*TN).expand(REG_TILES_PER_WAVE_M*TM, REG_TILES_PER_WAVE_N*TN)
+  c_regs = c_regs.after(c_regs.store(c_regs.after(k) + (A_col * B_row)).end(k).barrier().end(k_tile_range))
+
+  # store back to c
+  c_store = c.rearrange("(wpb_m rt_m lpw_m t_m) (wpb_n rt_n lpw_n t_n) -> (wpb_m wpb_n lpw_m lpw_n) (rt_m t_m) (rt_n t_n)", **consts)
+  return c_store[tid].store(c_regs).end(tid)
+
+def amd_copy_matmul(c:UOp, a:UOp, b:UOp) -> UOp:
+  block_id_n = UOp.range(N // BLOCK_N, 0, AxisType.GLOBAL)
+  block_id_m = UOp.range(M // BLOCK_M, 1, AxisType.GLOBAL)
+
+  # index the output with the globals
+  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :]
+  a = a.T.reshape(K, M // BLOCK_M, BLOCK_M)[:, block_id_m, :]
+  b = b.reshape(K, N // BLOCK_N, BLOCK_N)[:, block_id_n, :]
+  return block_128x128_gemm(c, a, b).end(block_id_n, block_id_m).sink(arg=KernelInfo(opts_to_apply=()))
+
+if __name__ == "__main__":
+  from amd_uop_matmul import eval_custom_matmul
+  eval_custom_matmul(amd_copy_matmul)

extra/gemm/amd_uop_matmul.py

@@ -1,98 +1,74 @@
-import numpy as np
-from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
+from tinygrad import Tensor, Context, GlobalCounters, dtypes
 from tinygrad.uop.ops import UOp, KernelInfo, sint, AxisType
-from tinygrad.engine.realize import ExecItem, get_runner
 from tinygrad.dtype import AddrSpace
-from tinygrad.helpers import getenv
+from tinygrad.helpers import DEBUG, getenv
 
 N = getenv("N", 4096)
-M = K = N
-run_count = getenv("CNT", 5)
+M = getenv("M", N)
+K = getenv("K", N)
+NUM_RUNS = getenv("CNT", 5)
 
 # ---------------------------
 # launch/config constants
 # ---------------------------
 
 WARP_SIZE = 32
-
-# Threadblock tile sizes (block-level tile of C that a block computes)
-BLOCK_N = 128   # columns of C (N-dim) per block
-BLOCK_M = 128   # rows of C (M-dim) per block
-BLOCK_K = 8     # K-slice per block iteration
-
-# Register tile sizes (per-thread accumulator tile of C)
-TN = 4     # columns per thread
-TM = 4     # rows per thread
+BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 8
+TM, TN = 4, 4
+LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8
+assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0
 
 is_kernel5 = getenv("K5", 0)
 THREADS_PER_BLOCK = 128 if is_kernel5 else 256
-assert THREADS_PER_BLOCK % BLOCK_N == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_N"
-assert THREADS_PER_BLOCK % BLOCK_K == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_K"
-assert (BLOCK_N * BLOCK_K) % THREADS_PER_BLOCK == 0
-assert (BLOCK_M * BLOCK_K) % THREADS_PER_BLOCK == 0
+WAVES_PER_BLOCK_N = 1 if is_kernel5 else 2
+WAVES_PER_BLOCK_M = THREADS_PER_BLOCK // WARP_SIZE // WAVES_PER_BLOCK_N
+REG_TILES_PER_WAVE_N = BLOCK_N // (WAVES_PER_BLOCK_N * LANES_PER_WAVE_N * TN)
+REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM)
 
-WARPS_PER_BLOCK = THREADS_PER_BLOCK // WARP_SIZE
-WAVE_TILE_N = 128 if is_kernel5 else 64
-WAVE_TILE_M = BLOCK_N * BLOCK_M // WARPS_PER_BLOCK // WAVE_TILE_N
-assert BLOCK_N % WAVE_TILE_N == 0, "BN must be a multiple of WN"
-assert BLOCK_M % WAVE_TILE_M == 0, "BM must be a multiple of WM"
-WAVES_IN_BLOCK_X = BLOCK_N // WAVE_TILE_N
-WAVES_IN_BLOCK_Y = BLOCK_M // WAVE_TILE_M
-assert WAVES_IN_BLOCK_X * WAVES_IN_BLOCK_Y == WARPS_PER_BLOCK, "wave grid must match warps/block"
-
-LANES_PER_WAVE_X = 8
-LANES_PER_WAVE_Y = 4
-ITERS_PER_WAVE_N = WAVE_TILE_N // (LANES_PER_WAVE_X * TN)
-ITERS_PER_WAVE_M = WAVE_TILE_M // (LANES_PER_WAVE_Y * TM)
-assert WAVE_TILE_N % (LANES_PER_WAVE_X * TN) == 0, "WAVE_TILE_N must be divisible by LANES_PER_WAVE_X*TN"
-assert WAVE_TILE_M % (LANES_PER_WAVE_Y * TM) == 0, "WAVE_TILE_M must be divisible by LANES_PER_WAVE_Y*TM"
+assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong"
+assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong"
 
 def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.LOOP): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
-def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=False):
+def copy(dest:UOp, src:UOp, rng:int, upcast=False):
   assert dest.shape == src.shape
   rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.LOOP)
-  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
-  return dest.after(copy) if set else copy
+  return dest[*rngs].store(src[*rngs]).end(*rngs)
 
-def hand_spec_kernel3():
+def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
   # ---------------------------
-  # block indices & placeholders
+  # block indices
   # ---------------------------
-  blockIdx_x = UOp.special(N // BLOCK_N, "gidx0")
-  blockIdx_y = UOp.special(N // BLOCK_M, "gidx1")
-
-  a = UOp.placeholder((N, N), dtypes.float, slot=1)
-  b = UOp.placeholder((N, N), dtypes.float, slot=2)
-  c = UOp.placeholder((N, N), dtypes.float, slot=0)
+  block_id_n = UOp.special(N // BLOCK_N, "gidx0")
+  block_id_m = UOp.special(M // BLOCK_M, "gidx1")
 
   # index the output with the globals
-  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[blockIdx_y, :, blockIdx_x, :]
+  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :]
 
   # open the main reduction range
-  k_tile_range = UOp.range(N // BLOCK_K, 0, AxisType.REDUCE)
-  a = a.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_K, BLOCK_K)[blockIdx_y, :, k_tile_range, :]
-  b = b.reshape(N // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, blockIdx_x, :]
+  k_tile_range = UOp.range(K // BLOCK_K, 0, AxisType.REDUCE)
+  a = a.reshape(M // BLOCK_M, BLOCK_M, K // BLOCK_K, BLOCK_K)[block_id_m, :, k_tile_range, :]
+  b = b.reshape(K // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, block_id_n, :]
 
   # globals are no longer used, they are already in the indexes
-  del blockIdx_y, blockIdx_x
+  del block_id_m, block_id_n
 
   # ---------------------------
-  # GLOBAL -> LOCAL (As, Bs)
+  # GLOBAL -> LOCAL (A_local, B_local)
   # ---------------------------
   tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
 
   # A: read BM x BK tiles (permute on store into locals)
-  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
-  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
-  As_store = copy(As.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)
+  BM_A_local_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
+  A_local = UOp.placeholder((BLOCK_K, BM_A_local_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
+  A_local_store = copy(A_local.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)
 
   # B: read BK x BN tiles
-  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
-  Bs_store = copy(Bs.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)
+  B_local = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
+  B_local_store = copy(B_local.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)
 
   # TODO: can we automate barrier?
-  barrier = UOp.barrier(As_store, Bs_store)
-  As, Bs = As.after(barrier), Bs.after(barrier)
+  barrier = UOp.barrier(A_local_store, B_local_store)
+  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
 
   # open inner k range
   k = UOp.range(BLOCK_K, 3, AxisType.REDUCE)
@@ -100,31 +76,30 @@ def hand_spec_kernel3():
   # ---------------------------
   # LOCAL -> REG (per-wave tiles)
   # ---------------------------
-  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
-  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
-  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y
+  warp, lane = tid // WARP_SIZE, tid % WARP_SIZE
+  waveIdx, waveIdy = warp % WAVES_PER_BLOCK_N, warp // WAVES_PER_BLOCK_N
+  laneIdx, laneIdy = lane % LANES_PER_WAVE_N, lane // LANES_PER_WAVE_N
+  assert waveIdy.vmax+1 == WAVES_PER_BLOCK_M and laneIdy.vmax+1 == LANES_PER_WAVE_M
 
-  laneIdx = (tid % WARP_SIZE) % LANES_PER_WAVE_X
-  laneIdy = (tid % WARP_SIZE) // LANES_PER_WAVE_X
-  assert laneIdy.vmax+1 == LANES_PER_WAVE_Y
+  A_col = UOp.placeholder((REG_TILES_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
+  A_local_slice = A_local[k, :].reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM)[waveIdy, :, laneIdy, :]
+  A_col = A_col.after(copy(A_col, A_local_slice, 300, upcast=True))
 
-  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  A_col = copy(A_col, As[k, :].reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)[waveIdy, :, laneIdy, :], 300, set=True, upcast=True)
-
-  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-  B_row = copy(B_row, Bs[k, :].reshape(WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)[waveIdx, :, laneIdx, :], 400, set=True, upcast=True)
+  B_row = UOp.placeholder((REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
+  B_local_slice = B_local[k, :].reshape(WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)[waveIdx, :, laneIdx, :]
+  B_row = B_row.after(copy(B_row, B_local_slice, 400, upcast=True))
 
   # ---------------------------
   # FMA: c_regs += A_col * B_row
   # ---------------------------
-  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  c_regs = UOp.placeholder((REG_TILES_PER_WAVE_M, TM, REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
   i = UOp.range(c_regs.size, 16)
   c_regs = c_regs.after(c_regs.flatten()[i].store(0.0).end(i))
 
   # TODO: why don't these work as upcast?
   # why if the ranges merge is it slow?!? (if you change the order on end, they will merge. big slowdown on METAL)
-  iterWaveM, yt, iterWaveN, xt = rngs = rngs_for_shape(c_regs.shape, 500)
-  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)
+  iter_m, t_m, iter_n, t_n = rngs = rngs_for_shape(c_regs.shape, 500)
+  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iter_m, t_m] * B_row[iter_n, t_n]).end(iter_m, iter_n, t_m, t_n)
 
   # Close k, sync, and close K tiles
   sink = sink.end(k).barrier().end(k_tile_range)
@@ -132,38 +107,37 @@ def hand_spec_kernel3():
   # ---------------------------
   # REG -> GLOBAL (epilogue)
   # ---------------------------
-  c = c.reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
-                WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
+  c = c.reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM,
+                WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)
   c = c[waveIdy, :, laneIdy, :,
         waveIdx, :, laneIdx, :]
   sink = copy(c, c_regs.after(sink), rng=600)
 
   return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()
 
-def test_matmul(sink:UOp, dtype=dtypes.float32, N=N):
-  rng = np.random.default_rng()
-  a = Tensor(rng.random((N, N), dtype=np.float32)-0.5, dtype=dtype)
-  b = Tensor(rng.random((N, N), dtype=np.float32)-0.5, dtype=dtype)
-  hc = Tensor.empty(N, N, dtype=dtype)
-  Tensor.realize(a, b, hc)
-
-  ei = ExecItem(sink, [t.uop.buffer for t in [hc, a, b]], prg=get_runner(Device.DEFAULT, sink))
+def eval_custom_matmul(fxn):
+  a = Tensor.randn(M, K, dtype=dtypes.float)
+  b = Tensor.randn(K, N, dtype=dtypes.float)
+  c = Tensor.empty(M, N, dtype=dtypes.float)
+  with Context(DEBUG=0): Tensor.realize(a, b)
 
   ets = []
-  with Context(DEBUG=2):
-    for _ in range(run_count):
-      ets.append(ei.run(wait=True))
-  print(f"REAL TFLOPS {N * N * N * 2 / min(ets) * 1e-12:.2f}")
+  with Context(DEBUG=max(2, DEBUG.value)):
+    for _ in range(NUM_RUNS):
+      GlobalCounters.reset()
+      tst = Tensor.custom_kernel(c, a, b, fxn=fxn)[0].realize()
+      ets.append(GlobalCounters.time_sum_s)
+  print(f"REAL TFLOPS {M * N * K * 2 / min(ets) * 1e-12:.2f}")
 
   if getenv("VERIFY", 1):
     GlobalCounters.reset()
     with Context(DEBUG=2):
       tc = (a @ b).realize()
     with Context(DEBUG=0):
-      err = (hc - tc).square().mean().item()
+      err = (tc - tst).square().mean().item()
     print(f"mean squared error {err}")
     if err > 1e-06:
       raise RuntimeError("matmul is wrong!")
 
 if __name__ == "__main__":
-  test_matmul(hand_spec_kernel3(), N=N)
+  eval_custom_matmul(hand_spec_kernel3)

extra/gemm/asm/cdna/gemm.py

@@ -1,122 +0,0 @@
-import atexit, functools
-from tinygrad import Tensor, Device, dtypes
-from tinygrad.dtype import AddrSpace
-from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
-from tinygrad.renderer import Estimates
-from tinygrad.helpers import getenv, all_same, DEBUG
-from extra.gemm.asm.cdna.asm import build_kernel, TILE_M, TILE_N, TILE_K, NUM_WG
-
-# ** CDNA4 assembly gemm
-
-WORKGROUP_SIZE = 256
-
-@functools.cache
-def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
-  batch, M, K = A.shape
-  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
-  assert K == K2
-  lidx = UOp.special(WORKGROUP_SIZE, "lidx0")
-  gidx = UOp.special(NUM_WG, "gidx0")
-  insts = build_kernel(batch, M, N, K, A.dtype.base)
-  lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=133_120, addrspace=AddrSpace.LOCAL), (), 'lds')
-  sink = UOp.sink(C.base, A.base, B.base, lds, lidx, gidx,
-                  arg=KernelInfo(name=f"gemm_{batch}_{M}_{N}_{K}", estimates=Estimates(ops=2*batch*M*N*K, mem=(batch*M*K + K*N + batch*M*N)*2)))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname),
-                                UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
-
-counters = {"used":0, "todos":[]}
-def todo(msg:str) -> bool: counters["todos"].append(msg); return False
-def _asm_gemm_report():
-  print(f'asm_gemm: {counters["used"]} used, {len(counters["todos"])} not used')
-  if DEBUG >= 2 and counters["todos"]:
-    from collections import Counter
-    for msg, cnt in Counter(counters["todos"]).most_common(): print(f'  {cnt:3d}x {msg}')
-atexit.register(_asm_gemm_report)
-
-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}")
-  batch, M, K = (1, *a.shape) if a.ndim == 2 else a.shape
-  N = b.shape[1]
-  if isinstance(a.device, tuple):
-    if a.ndim == 2 and a.uop.axis == 0 and b.uop.axis is None: M //= len(a.device)
-    elif a.ndim == 2 and a.uop.axis == 1 and b.uop.axis == 0: K //= len(a.device)
-    elif a.ndim == 2 and a.uop.axis is None and b.uop.axis == 1: N //= len(a.device)
-    elif a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None: batch //= len(a.device)
-    elif a.ndim == 3 and a.uop.axis is None and b.uop.axis == 1: N //= len(a.device)
-    elif a.ndim == 3 and a.uop.axis == 2 and b.uop.axis == 0: K //= 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 (M % TILE_M != 0 or N % TILE_N != 0 or K % TILE_K != 0) and arch == "gfx950":
-    return todo(f"GEMM shape ({M},{N},{K}) not a multiple of ({TILE_M},{TILE_N},{TILE_K})")
-  return True
-
-# ** UOp gemm to test Tensor.custom_kernel multi and backward correctness on non cdna4
-# note: this can be removed after we have GEMM on mixins
-
-def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
-  M, K = A.shape[0]*A.shape[1], A.shape[2]
-  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
-  assert K == K2
-  m = UOp.range(M, 1, AxisType.LOOP)
-  n = UOp.range(N, 2, AxisType.LOOP)
-  k = UOp.range(K, 0, AxisType.REDUCE)
-  mul = (A.index((m*UOp.const(dtypes.index, K)+k))*B.index((k*UOp.const(dtypes.index, N)+n))).cast(dtypes.float32)
-  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype.base)
-  store = C.index((m*UOp.const(dtypes.index, N)+n), ptr=True).store(red).end(m, n)
-  return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))
-
-# ** backward gemm, might use the asm gemm
-
-def custom_gemm_bw(gradient:UOp, kernel:UOp):
-  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
-  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
-
-def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
-  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
-  counters["used"] += 1
-  unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
-  if unfold_batch:
-    orig_batch = a.shape[0]
-    a = a.reshape(a.shape[0]*a.shape[1], a.shape[2])
-  squeeze = a.ndim == 2
-  if squeeze: a = a.unsqueeze(0)
-
-  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 (m_sharded:=is_multi and a.uop.axis == 1): M //= len(a.device)
-  n_sharded = is_multi and b.uop.axis == 1
-
-  if is_multi:
-    if n_sharded:
-      out = Tensor(Tensor.empty(batch, M, N//len(a.device), dtype=a.dtype, device=a.device).uop.multi(2), device=a.device)
-    elif m_sharded:
-      out = Tensor(Tensor.empty(batch, M, N, dtype=a.dtype, device=a.device).uop.multi(1), device=a.device)
-    else:
-      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)
-
-  renderer = Device[a.device[0] if is_multi else a.device].renderer
-  dname, arch = renderer.device, getattr(renderer, "arch", "")
-  if arch.startswith("gfx950") and getenv("USE_ASM", 1):
-    out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname), 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)
-  out = out.squeeze(0) if squeeze else out
-  if unfold_batch: out = out.reshape(orig_batch, -1, out.shape[-1])
-  return out

extra/gemm/asm/rdna3/template.s

@@ -1,76 +0,0 @@
-.text
-.section	.text.
-.global	gemm
-.p2align	8
-.type	gemm,@function
-
-gemm:
-INSTRUCTIONS
-
-.section .rodata,"a",@progbits
-.p2align 6, 0x0
-.amdhsa_kernel gemm
-  # basic memory requirements
-  .amdhsa_group_segment_fixed_size 30336
-  .amdhsa_private_segment_fixed_size 0
-  .amdhsa_kernarg_size 32
-  # register usage (RSRC1)
-  .amdhsa_next_free_vgpr 256
-  .amdhsa_next_free_sgpr 100
-  # workgroup / workitem IDs (RSRC2)
-  .amdhsa_system_sgpr_workgroup_id_x 1
-  .amdhsa_system_sgpr_workgroup_id_y 1
-  .amdhsa_system_sgpr_workgroup_id_z 1
-  # user SGPRs: kernarg ptr in s[0:1]
-  .amdhsa_user_sgpr_kernarg_segment_ptr 1
-  .amdhsa_user_sgpr_count 2
-  # gfx10+ / gfx11 specifics (RSRC1[29..31])
-  .amdhsa_wavefront_size32 1
-  .amdhsa_workgroup_processor_mode 1
-  .amdhsa_memory_ordered 1
-  .amdhsa_forward_progress 1
-  # misc for gfx11
-  .amdhsa_dx10_clamp 1
-  .amdhsa_ieee_mode 1
-  .amdhsa_uses_dynamic_stack 0
-.end_amdhsa_kernel
-
-.amdgpu_metadata
----
-amdhsa.kernels:
-  - .args:
-      - .address_space:  generic
-        .name:           C
-        .offset:         0
-        .size:           8
-        .value_kind:     global_buffer
-        .value_type:     f16
-      - .address_space:  generic
-        .name:           A
-        .offset:         8
-        .size:           8
-        .value_kind:     global_buffer
-        .value_type:     f16
-      - .address_space:  generic
-        .name:           B
-        .offset:         16
-        .size:           8
-        .value_kind:     global_buffer
-        .value_type:     f16
-    .group_segment_fixed_size: 30336
-    .kernarg_segment_align: 8
-    .kernarg_segment_size: 32
-    .max_flat_workgroup_size: 128
-    .name:           gemm
-    .private_segment_fixed_size: 0
-    .sgpr_count:     70
-    .sgpr_spill_count: 0
-    .symbol:         gemm.kd
-    .vgpr_count:     256
-    .vgpr_spill_count: 0
-    .wavefront_size: 32
-amdhsa.version:
-  - 1
-  - 1
-...
-.end_amdgpu_metadata

extra/gemm/asm/rdna3/test.py

@@ -1,30 +0,0 @@
-import math, pathlib
-
-from tinygrad import Device, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo
-
-from extra.gemm.amd_uop_matmul import test_matmul
-
-N = 4096
-TN = 96
-THREADS_PER_WG = 128
-NUM_WG = math.ceil(N / TN) * math.ceil(N / TN)
-
-dname:str = Device.DEFAULT
-template:str = (pathlib.Path(__file__).parent/"template.s").read_text()
-
-def asm_kernel() -> UOp:
-  lidx = UOp.special(THREADS_PER_WG, "lidx0")
-  gidx = UOp.special(NUM_WG, "gidx0")
-
-  a = UOp.placeholder((N*N,), dtypes.half, slot=1)
-  b = UOp.placeholder((N*N,), dtypes.half, slot=2)
-  c = UOp.placeholder((N*N,), dtypes.half, slot=0)
-
-  src = template.replace("INSTRUCTIONS", (pathlib.Path(__file__).parent/"gemm.s").read_text())
-
-  sink = UOp.sink(a, b, c, lidx, gidx, arg=KernelInfo(name="gemm"))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src)))
-
-if __name__ == "__main__":
-  test_matmul(asm_kernel(), dtype=dtypes.half, N=N)

extra/gemm/asm/unpack_kd.py

@@ -1,179 +0,0 @@
-# unpack the complete kernel descriptor of an amdgpu ELF
-# https://rocm.docs.amd.com/projects/llvm-project/en/latest/LLVM/llvm/html/AMDGPUUsage.html#code-object-v3-kernel-descriptor
-import struct, pathlib, sys
-from tinygrad.runtime.support.elf import elf_loader
-
-def bits(x, lo, hi): return (x >> lo) & ((1 << (hi - lo + 1)) - 1)
-def assert_zero(x, lo, hi): assert bits(x, lo, hi) == 0
-
-with open(sys.argv[1], "rb") as f:
-  lib = f.read()
-
-image, sections, relocs = elf_loader(lib)
-rodata_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".rodata"))
-
-# rodata is exactly 64 bytes
-kd = image[rodata_entry:rodata_entry+64]
-desc = int.from_bytes(kd, byteorder="little")
-
-group_segment_fixed_size = bits(desc, 0, 31)
-private_segment_fixed_size = bits(desc, 32, 63)
-kernarg_size = bits(desc, 64, 95)
-reserved_127_96 = bits(desc, 96, 127)
-assert reserved_127_96 == 0
-
-print("GROUP_SEGMENT_FIXED_SIZE:", group_segment_fixed_size)
-print("PRIVATE_SEGMENT_FIXED_SIZE:", private_segment_fixed_size)
-print("KERNARG_SIZE:", kernarg_size)
-print("RESERVED 127:96:", reserved_127_96)
-
-entry_off = bits(desc, 128, 191)
-
-# sign-extend manually if needed
-if entry_off & (1 << 63):
-  entry_off -= 1 << 64
-
-print("KERNEL_CODE_ENTRY_BYTE_OFFSET:", entry_off)
-
-kd_addr = 0x1840
-entry_addr = kd_addr + entry_off
-
-print("Computed entry address: 0x%016x" % entry_addr)
-print("256B aligned:", entry_addr % 256 == 0)
-
-pgm_rsrc3 = bits(desc, 352, 383)
-pgm_rsrc1 = bits(desc, 384, 415)
-pgm_rsrc2 = bits(desc, 416, 447)
-
-print("COMPUTE_PGM_RSRC3: 0x%08x" % pgm_rsrc3)
-print("COMPUTE_PGM_RSRC1: 0x%08x" % pgm_rsrc1)
-print("COMPUTE_PGM_RSRC2: 0x%08x" % pgm_rsrc2)
-
-# rsrc 3 (gfx950)
-
-accum_offset_raw = bits(pgm_rsrc3, 0, 5)
-assert_zero(pgm_rsrc3, 6, 15)
-tg_split = bits(pgm_rsrc3, 16, 16)
-accum_offset_vgprs = (accum_offset_raw + 1) * 4
-print("RSRC3.ACCUM_OFFSET (AccVGPR index):", accum_offset_vgprs)
-print("RSRC3.TG_SPLIT:", tg_split)
-
-# rsrc 1
-
-vgpr_gran = bits(pgm_rsrc1, 0, 5)
-sgpr_gran = bits(pgm_rsrc1, 6, 9)
-assert_zero(pgm_rsrc1, 27, 28)
-
-# NOTE: this is vgprs + agprs
-vgprs_used = (vgpr_gran + 1) * 8
-assert 0 <= vgprs_used <= 512
-
-k = sgpr_gran // 2
-sgprs_used = (k + 1) * 16
-
-print("RSRC1.VGPRS:", vgprs_used)
-print("RSRC1.SGPRS:", sgprs_used)
-
-assert_zero(pgm_rsrc1, 10, 11)
-
-float_round_mode_32 = bits(pgm_rsrc1, 12, 13)
-float_round_mode_16_64 = bits(pgm_rsrc1, 15, 14)
-float_denorm_mode_32 = bits(pgm_rsrc1, 16, 17)
-float_denorm_mode_16_64 = bits(pgm_rsrc1, 18, 19)
-
-priv = bits(pgm_rsrc1, 20, 20)
-assert priv == 0
-enable_dx10_clamp_wg_rr_en = bits(pgm_rsrc1, 21, 21)
-debug_mode = bits(pgm_rsrc1, 22, 22)
-enable_ieee_mode = bits(pgm_rsrc1, 23, 23)
-bulky = bits(pgm_rsrc1, 24, 24)
-assert bulky == 0
-cdbg_user = bits(pgm_rsrc1, 25, 25)
-assert cdbg_user == 0
-fp16_ovfl = bits(pgm_rsrc1, 26, 26)
-assert_zero(pgm_rsrc1, 27, 28)  # reserved
-assert_zero(pgm_rsrc1, 29, 29)  # WGP_MODE (reserved on gfx9)
-assert_zero(pgm_rsrc1, 30, 30)  # MEM_ORDERED (reserved on gfx9)
-assert_zero(pgm_rsrc1, 31, 31)  # FWD_PROGRESS (reserved on gfx9)
-
-# rsrc 2
-
-enable_private_segment = bits(pgm_rsrc2, 0, 0)  # SCRATCH_EN
-user_sgpr_count = bits(pgm_rsrc2, 1, 5)         # USER_SGPR
-enable_trap_handler = bits(pgm_rsrc2, 6, 6)     # TRAP_PRESENT (must be 0 here)
-assert enable_trap_handler == 0
-
-enable_sgpr_workgroup_id_x = bits(pgm_rsrc2, 7, 7)
-enable_sgpr_workgroup_id_y = bits(pgm_rsrc2, 8, 8)
-enable_sgpr_workgroup_id_z = bits(pgm_rsrc2, 9, 9)
-enable_sgpr_workgroup_info = bits(pgm_rsrc2, 10, 10)
-
-enable_vgpr_workitem_id = bits(pgm_rsrc2, 11, 12)  # TIDIG_CMP_CNT enum (0..3)
-
-enable_exception_address_watch = bits(pgm_rsrc2, 13, 13)
-assert enable_exception_address_watch == 0
-enable_exception_memory = bits(pgm_rsrc2, 14, 14)
-assert enable_exception_memory == 0
-
-granulated_lds_size = bits(pgm_rsrc2, 15, 23)
-assert granulated_lds_size == 0  # spec: must be 0; CP uses dispatch packet rounding
-
-enable_exception_fp_invalid = bits(pgm_rsrc2, 24, 24)
-enable_exception_fp_denorm_src = bits(pgm_rsrc2, 25, 25)
-enable_exception_fp_div0 = bits(pgm_rsrc2, 26, 26)
-enable_exception_fp_overflow = bits(pgm_rsrc2, 27, 27)
-enable_exception_fp_underflow = bits(pgm_rsrc2, 28, 28)
-enable_exception_fp_inexact = bits(pgm_rsrc2, 29, 29)
-enable_exception_int_div0 = bits(pgm_rsrc2, 30, 30)
-
-assert_zero(pgm_rsrc2, 31, 31)
-
-print("RSRC2.ENABLE_PRIVATE_SEGMENT:", enable_private_segment)
-print("RSRC2.USER_SGPR_COUNT:", user_sgpr_count)
-print("RSRC2.ENABLE_SGPR_WORKGROUP_ID_X:", enable_sgpr_workgroup_id_x)
-print("RSRC2.ENABLE_SGPR_WORKGROUP_ID_Y:", enable_sgpr_workgroup_id_y)
-print("RSRC2.ENABLE_SGPR_WORKGROUP_ID_Z:", enable_sgpr_workgroup_id_z)
-print("RSRC2.ENABLE_SGPR_WORKGROUP_INFO:", enable_sgpr_workgroup_info)
-print("RSRC2.ENABLE_VGPR_WORKITEM_ID (enum):", enable_vgpr_workitem_id)
-
-print("RSRC2.EXC_FP_INVALID:", enable_exception_fp_invalid)
-print("RSRC2.EXC_FP_DENORM_SRC:", enable_exception_fp_denorm_src)
-print("RSRC2.EXC_FP_DIV0:", enable_exception_fp_div0)
-print("RSRC2.EXC_FP_OVERFLOW:", enable_exception_fp_overflow)
-print("RSRC2.EXC_FP_UNDERFLOW:", enable_exception_fp_underflow)
-print("RSRC2.EXC_FP_INEXACT:", enable_exception_fp_inexact)
-print("RSRC2.EXC_INT_DIV0:", enable_exception_int_div0)
-
-# user sgprs
-
-enable_sgpr_private_segment_buffer = bits(desc, 448, 448)
-enable_sgpr_dispatch_ptr = bits(desc, 449, 449)
-enable_sgpr_queue_ptr = bits(desc, 450, 450)
-enable_sgpr_kernarg_segment_ptr = bits(desc, 451, 451)
-enable_sgpr_dispatch_id = bits(desc, 452, 452)
-enable_sgpr_flat_scratch_init = bits(desc, 453, 453)
-enable_sgpr_private_segment_size = bits(desc, 454, 454)
-
-assert_zero(desc, 455, 457)
-
-print("DESC.ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER:", enable_sgpr_private_segment_buffer)
-print("DESC.ENABLE_SGPR_DISPATCH_PTR:", enable_sgpr_dispatch_ptr)
-print("DESC.ENABLE_SGPR_QUEUE_PTR:", enable_sgpr_queue_ptr)
-print("DESC.ENABLE_SGPR_KERNARG_SEGMENT_PTR:", enable_sgpr_kernarg_segment_ptr)
-print("DESC.ENABLE_SGPR_DISPATCH_ID:", enable_sgpr_dispatch_id)
-print("DESC.ENABLE_SGPR_FLAT_SCRATCH_INIT:", enable_sgpr_flat_scratch_init)
-print("DESC.ENABLE_SGPR_PRIVATE_SEGMENT_SIZE:", enable_sgpr_private_segment_size)
-
-assert_zero(desc, 458, 459)
-
-uses_dynamic_stack = bits(desc, 459, 460)
-print("DESC.USES_DYNAMIC_STACK:", uses_dynamic_stack)
-
-# gfx950 only
-assert_zero(desc, 460, 463)
-kernarg_preload_spec_length = bits(desc, 464, 470)
-print("DESC.KERNARG_PRELOAD_SPEC_LENGTH:", kernarg_preload_spec_length)
-kernarg_preload_spec_offset = bits(desc, 471, 479)
-print("DESC.KERNARG_PRELOAD_SPEC_OFFSET:", kernarg_preload_spec_offset)
-
-assert_zero(desc, 480, 511)

extra/gemm/cdna_asm_gemm.py

@@ -1,5 +1,14 @@
+import atexit, functools
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.dtype import AddrSpace
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from tinygrad.renderer import Estimates
+from tinygrad.helpers import getenv, all_same, DEBUG
 from tinygrad.runtime.autogen.amd.cdna.ins import *
-from tinygrad.dtype import dtypes
+
+# ** CDNA4 assembly gemm
+
+WORKGROUP_SIZE = 256
 
 # M0 is encoded with 124 (NULL in RDNA) in CDNA
 M0 = NULL
@@ -2597,3 +2606,117 @@ def build_kernel(batch, M, N, K, dtype):
   k.label('KernelEnd')
   k.emit(s_endpgm())
   return k.finalize()
+
+# ** ASM_GEMM custom kernel
+
+@functools.cache
+def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
+  batch, M, K = A.shape
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  lidx = UOp.special(WORKGROUP_SIZE, "lidx0")
+  gidx = UOp.special(NUM_WG, "gidx0")
+  insts = build_kernel(batch, M, N, K, A.dtype.base)
+  lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=133_120, addrspace=AddrSpace.LOCAL), (), 'lds')
+  sink = UOp.sink(C.base, A.base, B.base, lds, lidx, gidx,
+                  arg=KernelInfo(name=f"gemm_{batch}_{M}_{N}_{K}", estimates=Estimates(ops=2*batch*M*N*K, mem=(batch*M*K + K*N + batch*M*N)*2)))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname),
+                                UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+
+counters = {"used":0, "todos":[]}
+def todo(msg:str) -> bool: counters["todos"].append(msg); return False
+def _asm_gemm_report():
+  print(f'asm_gemm: {counters["used"]} used, {len(counters["todos"])} not used')
+  if DEBUG >= 2 and counters["todos"]:
+    from collections import Counter
+    for msg, cnt in Counter(counters["todos"]).most_common(): print(f'  {cnt:3d}x {msg}')
+atexit.register(_asm_gemm_report)
+
+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}")
+  batch, M, K = (1, *a.shape) if a.ndim == 2 else a.shape
+  N = b.shape[1]
+  if isinstance(a.device, tuple):
+    if a.ndim == 2 and a.uop.axis == 0 and b.uop.axis is None: M //= len(a.device)
+    elif a.ndim == 2 and a.uop.axis == 1 and b.uop.axis == 0: K //= len(a.device)
+    elif a.ndim == 2 and a.uop.axis is None and b.uop.axis == 1: N //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None: batch //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis is None and b.uop.axis == 1: N //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis == 2 and b.uop.axis == 0: K //= 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 (M % TILE_M != 0 or N % TILE_N != 0 or K % TILE_K != 0) and arch == "gfx950":
+    return todo(f"GEMM shape ({M},{N},{K}) not a multiple of ({TILE_M},{TILE_N},{TILE_K})")
+  return True
+
+# ** UOp gemm to test Tensor.custom_kernel multi and backward correctness on non cdna4
+# note: this can be removed after we have GEMM on mixins
+
+def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
+  M, K = A.shape[0]*A.shape[1], A.shape[2]
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  m = UOp.range(M, 1, AxisType.LOOP)
+  n = UOp.range(N, 2, AxisType.LOOP)
+  k = UOp.range(K, 0, AxisType.REDUCE)
+  mul = (A.flatten().index((m*UOp.const(dtypes.index, K)+k))*
+         B.flatten().index((k*UOp.const(dtypes.index, N)+n))).cast(dtypes.float32)
+  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype.base)
+  store = C.flatten().index((m*UOp.const(dtypes.index, N)+n), ptr=True).store(red).end(m, n)
+  return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))
+
+# ** backward gemm, might use the asm gemm
+
+def custom_gemm_bw(gradient:UOp, kernel:UOp):
+  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
+  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
+
+def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
+  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
+  counters["used"] += 1
+  unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
+  if unfold_batch:
+    orig_batch = a.shape[0]
+    a = a.reshape(a.shape[0]*a.shape[1], a.shape[2])
+  squeeze = a.ndim == 2
+  if squeeze: a = a.unsqueeze(0)
+
+  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 (m_sharded:=is_multi and a.uop.axis == 1): M //= len(a.device)
+  n_sharded = is_multi and b.uop.axis == 1
+
+  if is_multi:
+    if n_sharded:
+      out = Tensor(Tensor.empty(batch, M, N//len(a.device), dtype=a.dtype, device=a.device).uop.multi(2), device=a.device)
+    elif m_sharded:
+      out = Tensor(Tensor.empty(batch, M, N, dtype=a.dtype, device=a.device).uop.multi(1), device=a.device)
+    else:
+      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)
+
+  renderer = Device[a.device[0] if is_multi else a.device].renderer
+  dname, arch = renderer.device, getattr(renderer, "arch", "")
+  if arch.startswith("gfx950") and getenv("USE_ASM", 1):
+    out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname), 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)
+  out = out.squeeze(0) if squeeze else out
+  if unfold_batch: out = out.reshape(orig_batch, -1, out.shape[-1])
+  return out

extra/sqtt/examples/discover_ops.py

@@ -1,148 +0,0 @@
-#!/usr/bin/env python3
-# Run all ALU and memory instructions in the ISA
-import functools, inspect
-from enum import Enum
-from tinygrad import Tensor, Device, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo, AddrSpace
-from tinygrad.renderer.amd.dsl import Inst, Reg, OPERANDS, SrcField, VGPRField, SGPRField, SSrcField, SBaseField, AlignedSGPRField, BitField
-from tinygrad.renderer.amd.dsl import FixedBitField, EnumBitField, s, v, NULL, VCC_LO
-from extra.gemm.amd_asm_matmul import Kernel
-
-# skip instructions that mutate wave state (PC, EXEC, allocations, signals)
-SKIP = {"S_SETPC_B64", "S_SWAPPC_B64", "S_RFE_B64", "S_BARRIER_SIGNAL_ISFIRST", "S_GET_BARRIER_STATE", "S_ALLOC_VGPR", "S_SLEEP_VAR", "S_GETPC_B64",
-        "S_SENDMSG_RTN_B32", "S_SENDMSG_RTN_B64"}
-# skip barriers, s_waits, wrap level atomics, and ray tracing (bvh)
-SKIP_SUBSTR = ["SAVEEXEC", "CMPX", "WREXEC", "MOVREL", "ATOMIC", "S_BUFFER_", "S_ATC_PROBE", "BARRIER", "S_WAITCNT", "BVH",
-               "DS_CMPSTORE_RTN", "DS_WRAP_RTN_B32", "DS_ORDERED_COUNT", "DS_GWS", "GS_REG", "GLOBAL_LOAD_LDS", "GLOBAL_STORE_BLOCK"]
-
-ALU_FORMATS = {"VOP1", "VOP1_LIT", "VOP1_SDST", "VOP2", "VOP2_LIT", "VOP3", "VOP3_SDST", "VOP3SD", "VOP3P", "VOP3P_MFMA", "VOP3PX2",
-               "VOPC", "SOP1", "SOP1_LIT", "SOP2", "SOP2_LIT", "SOPC", "SOPC_LIT", "SOPK", "SOPK_LIT", "VINTERP"}
-# intentionally not testing scratch memory ops
-MEM_FORMATS = {"VGLOBAL", "GLOBAL", "SMEM", "DS"}
-
-def should_skip(op:Enum) -> bool: return (name:=op.name) in SKIP or any(sub in name for sub in SKIP_SUBSTR)
-
-# ** named register assignments
-
-# ALU operands
-ALU_VGPR_STRIDE = 16          # v[0], v[16], v[32], ... per ALU operand slot
-ALU_SGPR_STRIDE = 4           # s[0], s[4], s[8], ... per ALU operand slot
-
-# memory address registers
-S_KERNARG_PTR = (0, 1)
-S_BUF_PTR = (2, 3)
-V_VADDR = (0, 1)
-V_DS_ADDR = 0
-
-# memory data registers
-MEM_VGPR_BASE = 32            # v[32], v[48], ... for vdst/vdata/vsrc
-MEM_VGPR_STRIDE = 16          # spacing between memory data vgpr slots
-MEM_SGPR_BASE = 8             # s[8], s[10], ... for SMEM sdata
-MEM_SGPR_STRIDE = 2           # spacing between memory data sgpr slots
-
-# ** create an ALU instruction based on the operands
-
-def create_alu_inst(op:Enum, builder:functools.partial[Inst]) -> Inst:
-  inst_cls, operands, slot = builder.func, OPERANDS[op], 0
-  kwargs:dict[str, Reg|int] = {}
-  for name, field in inst_cls._fields:
-    if isinstance(field, (FixedBitField, EnumBitField)): continue
-    nregs = max(1, operands[name][1] // 32) if name in operands else 1
-    is_sreg = name in operands and "SREG" in str(operands[name][2])
-    base_v, base_s = slot * ALU_VGPR_STRIDE, slot * ALU_SGPR_STRIDE
-    if name == "sdst" and isinstance(field, SGPRField): reg = VCC_LO
-    elif is_sreg and not isinstance(field, VGPRField): reg = VCC_LO
-    elif isinstance(field, VGPRField): reg = v[base_v:base_v+nregs-1] if nregs > 1 else v[base_v]
-    elif isinstance(field, SSrcField): reg = VCC_LO if nregs <= 2 else s[base_s:base_s+nregs-1] if nregs > 1 else s[base_s]
-    elif isinstance(field, SGPRField): reg = s[base_s:base_s+nregs-1] if nregs > 1 else s[base_s]
-    elif isinstance(field, SrcField): reg = v[base_v:base_v+nregs-1] if nregs > 1 else v[base_v]
-    else: reg = None
-    if reg is not None: kwargs[name] = reg; slot += 1
-    elif isinstance(field, BitField): kwargs[name] = field.default
-  return builder(**kwargs)
-
-# ** create a memory instruction with pre set address registers
-
-MEM_PRESET_REGS:dict[str, dict[str, Reg]] = {
-  "VGLOBAL":{"saddr":s[S_BUF_PTR[0]:S_BUF_PTR[1]], "vaddr":v[V_VADDR[0]:V_VADDR[1]]},
-  "GLOBAL":{"saddr":s[S_BUF_PTR[0]:S_BUF_PTR[1]], "addr":v[V_DS_ADDR]},  # addr is 32-bit offset when saddr is valid SGPR
-  "DS":{"addr":v[V_DS_ADDR]},
-  "SMEM":{"sbase":s[S_KERNARG_PTR[0]:S_KERNARG_PTR[1]], "soffset":NULL},
-}
-
-def create_mem_inst(op:Enum, builder:functools.partial[Inst]) -> Inst:
-  inst_cls, operands, field_map = builder.func, OPERANDS.get(op, {}), MEM_PRESET_REGS.get(builder.func.__name__, {})
-  kwargs:dict[str, Reg|int] = {}
-  vslot, sslot = 0, 0
-  for name, field in inst_cls._fields:
-    if isinstance(field, (FixedBitField, EnumBitField)): continue
-    if name in field_map:
-      kwargs[name] = field_map[name]
-      continue
-    nregs = max(1, operands[name][1] // 32) if name in operands else 1
-    if isinstance(field, VGPRField):
-      vi = MEM_VGPR_BASE + vslot * MEM_VGPR_STRIDE
-      kwargs[name] = v[vi:vi+nregs-1] if nregs > 1 else v[vi]
-      vslot += 1
-    elif isinstance(field, (SGPRField, AlignedSGPRField, SBaseField)):
-      si = MEM_SGPR_BASE + sslot * MEM_SGPR_STRIDE
-      kwargs[name] = s[si:si+nregs-1] if nregs > 1 else s[si]
-      sslot += 1
-    elif isinstance(field, BitField): kwargs[name] = field.default
-  return builder(**kwargs)
-
-# ** collect all memory and ALU instructions from the ISA autogen
-
-def collect_instructions() -> tuple[list[Inst], list[Inst], list[str]]:
-  op_map:dict[Enum, functools.partial[Inst]] = {}
-  for name, obj in inspect.getmembers(all_insts):
-    if isinstance(obj, functools.partial) and len(obj.args) == 1: op_map[obj.args[0]] = obj
-  alu_insts:list[Inst] = []
-  mem_insts:list[Inst] = []
-  skipped:list[str] = []
-  for op_enum, builder in op_map.items():
-    if should_skip(op_enum) or op_enum not in OPERANDS: skipped.append(op_enum.name); continue
-    fmt = builder.func.__name__
-    if fmt in ALU_FORMATS: alu_insts.append(create_alu_inst(op_enum, builder))
-    elif fmt in MEM_FORMATS: mem_insts.append(create_mem_inst(op_enum, builder))
-  return alu_insts, mem_insts, skipped
-
-def exec_insts(insts:list):
-  k = Kernel(arch)
-  # ** prologue for global memory
-  k.emit(s_load_b64(sdata=s[S_BUF_PTR[0]:S_BUF_PTR[1]], sbase=s[S_KERNARG_PTR[0]:S_KERNARG_PTR[1]], soffset=NULL))
-  k.waitcnt(lgkm=0)
-  k.emit(v_mov_b32_e32(v[V_VADDR[0]], 0))
-  k.emit(v_mov_b32_e32(v[V_VADDR[1]], 0))
-  # ** emit
-  for inst in insts: k.emit(inst)
-  k.emit(s_endpgm())
-  # ** run
-  NUM_THREADS, NUM_GRIDS, BUF_SIZE = 32, 1, 1024*1024
-  def fxn(A:UOp, B:UOp, C:UOp) -> UOp:
-    lidx, gidx = UOp.special(NUM_THREADS, "lidx0"), UOp.special(NUM_GRIDS, "gidx0")
-    lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=BUF_SIZE, addrspace=AddrSpace.LOCAL), (), "lds")
-    sink = UOp.sink(A.base, B.base, C.base, lds, lidx, gidx, arg=KernelInfo(name="discover_ops"))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple(UOp(Ops.INS, arg=x) for x in k.finalize()))))
-  A = Tensor.empty(BUF_SIZE, dtype=dtypes.uint8)
-  B = Tensor.empty(1, dtype=dtypes.uint8)
-  C = Tensor.empty(1, dtype=dtypes.uint8)
-  Tensor.custom_kernel(A, B, C, fxn=fxn)[0].realize()
-
-if __name__ == "__main__":
-  import sys
-  arch = Device[Device.DEFAULT].renderer.arch
-  if arch.startswith("gfx12"):
-    from tinygrad.runtime.autogen.amd.rdna4.ins import *
-    import tinygrad.runtime.autogen.amd.rdna4.ins as all_insts
-  elif arch.startswith("gfx11"):
-    from tinygrad.runtime.autogen.amd.rdna3.ins import *
-    import tinygrad.runtime.autogen.amd.rdna3.ins as all_insts
-    # these don"t exist in RDNA3, only RDNA3.5 and above
-    SKIP.update(["S_FMAAK_F32", "S_FMAMK_F32"])
-  else:
-    print(f"{arch} not supported yet")
-    sys.exit(0)
-  alu_insts, mem_insts, skipped = collect_instructions()
-  print(f"collected {len(alu_insts)} ALU + {len(mem_insts)} memory instructions ({len(skipped)} skipped)")
-  exec_insts(mem_insts+alu_insts)

extra/sqtt/examples/generate_examples.py

@@ -8,8 +8,8 @@ PROFILE_PATH = Path(temp("profile.pkl", append_user=True))
 EXAMPLES = {
   "empty":"test/backend/test_custom_kernel.py TestCustomKernel.test_empty",
   "plus":"test/test_tiny.py TestTiny.test_plus",
-  "gemm":"-c \"from tinygrad import Tensor; (Tensor.empty(N:=64, N)@Tensor.empty(N, N)).realize()\"",
-  "ops":"extra/sqtt/examples/discover_ops.py"
+  "gemm":"-c \"from tinygrad import Tensor; (Tensor.empty(N:=32, N)@Tensor.empty(N, N)).realize()\"",
+  "sync":"test/amd/test_custom_kernel.py TestCustomKernel.test_wave_sync",
 }
 
 if __name__ == "__main__":

extra/torch_backend/backend.py

@@ -590,7 +590,7 @@ tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
   "aten.repeat": lambda x,*repeats: Tensor.repeat(x,*repeats).contiguous(), # not a view
   "aten._softmax": lambda self,dim,half_to_float: self.softmax(dim),
   "aten._log_softmax": lambda self,dim,half_to_float: self.log_softmax(dim),
-  "aten.random_": lambda self: Tensor.randint(*self.shape, low=dtypes.min(self.dtype), high=dtypes.max(self.dtype), device=self.device, dtype=self.dtype),
+  "aten.random_": lambda self: Tensor.randint(*self.shape, low=self.dtype.min, high=self.dtype.max, device=self.device, dtype=self.dtype),
   "aten.random_.from": lambda self, from_, to: Tensor.randint(*self.shape, low=from_, high=to, device=self.device, dtype=self.dtype),
   "aten.uniform_": lambda self, low=0, high=1: Tensor.uniform(*self.shape, low=low, high=high, dtype=self.dtype),
   "aten.normal_": lambda self, mean=0, std=1: Tensor.normal(*self.shape, mean=mean, std=std, dtype=self.dtype),

extra/usbgpu/tbgpu/installer/Shared/server.c

@@ -17,6 +17,7 @@
 // Protocol
 
 enum {
+  CMD_PROBE = 0,          // probe devices, returns count
   CMD_MAP_BAR = 1,        // map PCI BAR, returns size
   CMD_MAP_SYSMEM_FD = 2,  // alloc DMA memory, returns fd via SCM_RIGHTS
   CMD_CFG_READ = 3,       // read PCI config space
@@ -24,11 +25,14 @@ enum {
   CMD_RESET = 5,          // reset device
   CMD_MMIO_READ = 6,      // bulk read from BAR
   CMD_MMIO_WRITE = 7,     // bulk write to BAR
+  CMD_MAP_SYSMEM = 8,     // map system memory
+  CMD_SYSMEM_READ = 9,    // bulk read from system memory
+  CMD_SYSMEM_WRITE = 10,  // bulk write to system memory
   RESP_OK = 0, RESP_ERR = 1,
 };
 
-typedef struct { uint8_t cmd, bar; uint64_t offset, size, value; } __attribute__((packed)) request_t;
-typedef struct { uint8_t status; uint64_t value, addr; } __attribute__((packed)) response_t;
+typedef struct { uint8_t cmd; uint32_t dev_id, bar; uint64_t arg0, arg1, arg2; } __attribute__((packed)) request_t;
+typedef struct { uint8_t status; uint64_t resp0, resp1; } __attribute__((packed)) response_t;
 
 // Constants and state
 
@@ -77,7 +81,7 @@ static int send_response(int fd, response_t *resp, int send_fd) {
 }
 
 static void send_error(int fd, const char *msg) {
-  response_t resp = {.status = RESP_ERR, .value = strlen(msg)};
+  response_t resp = {.status = RESP_ERR, .resp0 = strlen(msg)};
   send_response(fd, &resp, -1);
   send(fd, msg, strlen(msg), 0);
 }
@@ -116,8 +120,8 @@ static int dext_rpc(uint32_t sel, uint64_t *in, uint32_t in_cnt, uint64_t *out_v
 static int map_bar(uint32_t bar, response_t *resp) {
   if (bar >= MAX_BARS) return -1;
   if (!g_bars[bar].addr && IOConnectMapMemory64(g_conn, bar, mach_task_self(), &g_bars[bar].addr, &g_bars[bar].size, kIOMapAnywhere)) return -1;
-  resp->addr = g_bars[bar].addr;
-  resp->value = g_bars[bar].size;
+  resp->resp0 = g_bars[bar].addr;
+  resp->resp1 = g_bars[bar].size;
   return 0;
 }
 
@@ -148,7 +152,7 @@ static int map_sysmem_fd(uint64_t size, response_t *resp, int *out_fd) {
   strncpy(g_sysmem[idx].shm_name, shm_name, sizeof(g_sysmem[idx].shm_name));
   g_sysmem_count++;
 
-  *resp = (response_t){.addr = idx, .value = alloc_sz};
+  *resp = (response_t){.resp0 = alloc_sz, .resp1 = idx};
   *out_fd = fd;
   return 0;
 
@@ -203,19 +207,19 @@ static void handle_client(int fd) {
 
     case CMD_MAP_SYSMEM_FD: {
       int shm_fd = -1;
-      resp.status = map_sysmem_fd(req.size, &resp, &shm_fd) ? 1 : 0;
+      resp.status = map_sysmem_fd(req.arg0, &resp, &shm_fd) ? 1 : 0;
       send_response(fd, &resp, shm_fd);
       continue;
     }
 
     case CMD_CFG_READ: {
-      uint64_t in[2] = {req.offset, req.size};
-      resp.status = dext_rpc(0, in, 2, &resp.value) ? 1 : 0;
+      uint64_t in[2] = {req.arg0, req.arg1};
+      resp.status = dext_rpc(0, in, 2, &resp.resp0) ? 1 : 0;
       break;
     }
 
     case CMD_CFG_WRITE: {
-      uint64_t in[3] = {req.offset, req.size, req.value};
+      uint64_t in[3] = {req.arg0, req.arg1, req.arg2};
       resp.status = dext_rpc(1, in, 3, NULL) ? 1 : 0;
       break;
     }
@@ -225,17 +229,17 @@ static void handle_client(int fd) {
       break;
 
     case CMD_MMIO_READ:
-      if (validate_bar(req.bar, req.offset, req.size)) { resp.status = 1; break; }
-      mmio_copy(g_bulk_buf, (void*)(g_bars[req.bar].addr + req.offset), req.size);
-      resp.value = req.size;
+      if (validate_bar(req.bar, req.arg0, req.arg1)) { resp.status = 1; break; }
+      mmio_copy(g_bulk_buf, (void*)(g_bars[req.bar].addr + req.arg0), req.arg1);
+      resp.resp0 = req.arg1;
       send_response(fd, &resp, -1);
-      send(fd, g_bulk_buf, req.size, 0);
+      send(fd, g_bulk_buf, req.arg1, 0);
       continue;
 
     case CMD_MMIO_WRITE:
-      recvall(fd, g_bulk_buf, req.size);
-      if (!validate_bar(req.bar, req.offset, req.size))
-        mmio_copy((void*)(g_bars[req.bar].addr + req.offset), g_bulk_buf, req.size);
+      recvall(fd, g_bulk_buf, req.arg1);
+      if (!validate_bar(req.bar, req.arg0, req.arg1))
+        mmio_copy((void*)(g_bars[req.bar].addr + req.arg0), g_bulk_buf, req.arg1);
       continue;
 
     default:

extra/usbgpu/tbgpu/installer/notary_tool.sh

@@ -3,3 +3,7 @@ set -e
 
 ditto -c -k --keepParent ./build/Release/TinyGPU.app ./build/Release/TinyGPU.zip
 xcrun notarytool submit ./build/Release/TinyGPU.zip --keychain-profile "hgwJFhdheiIEy82nDN" --wait
+
+rm ./build/Release/TinyGPU.zip
+xcrun stapler staple ./build/Release/TinyGPU.app
+ditto -c -k --keepParent ./build/Release/TinyGPU.app ./build/Release/TinyGPU.zip

extra/viz/cli.py

@@ -48,11 +48,16 @@ def decode_profile(data:bytes) -> dict:
         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["linear"] = u("<B")[0]
       v["peak"] = u("<Q")[0]
       for _ in range(event_count):
-        alloc, ts, key = u("<BII")
-        if alloc: v["events"].append({"event":"alloc", "ts":ts, "key":key, "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
-        else: v["events"].append({"event":"free", "ts":ts, "key":key, "arg": {"users":[u("<IIIB") for _ in range(u("<I")[0])]}})
+        if v["linear"]:
+          ts, value = u("<IQ")
+          v["events"].append({"event":"freq", "ts":ts, "value":value})
+        else:
+          alloc, ts, key = u("<BII")
+          if alloc: v["events"].append({"event":"alloc", "ts":ts, "key":key, "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
+          else: v["events"].append({"event":"free", "ts":ts, "key":key, "arg": {"users":[u("<IIIB") for _ in range(u("<I")[0])]}})
   return {"dur":total_dur, "peak":global_peak, "layout":layout, "markers":markers}
 
 if __name__ == "__main__":

pyproject.toml

@@ -74,7 +74,7 @@ testing_minimal = [
   "hypothesis>=6.148.9",
   "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", "gguf"]
+testing_unit = ["tinygrad[testing_minimal]", "tqdm", "safetensors", "tabulate", "openai", "gguf>=0.18"]
 testing = [
   "tinygrad[testing_unit]",
   "pillow",

test/amd/helpers.py

@@ -5,8 +5,8 @@ from tinygrad.runtime.autogen import llvm
 from tinygrad.runtime.support.elf import elf_loader
 
 ARCH_TO_TARGET:dict[str, list[str]] = {
-  "rdna3":["gfx1100"],
-  "rdna4":["gfx1200"],
+  "rdna3":["gfx1100", "gfx1151"],
+  "rdna4":["gfx1200", "gfx1201"],
   "cdna":["gfx950", "gfx942"],
 }
 

test/amd/test_custom_kernel.py

@@ -1,10 +1,12 @@
 import unittest
+import functools
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.renderer import Estimates
-
 from tinygrad.runtime.autogen.amd.rdna3.ins import *
+from tinygrad.runtime.autogen.amd.rdna4.ins import s_barrier_wait, s_barrier_signal
 from tinygrad.renderer.amd.dsl import s, v
+from test.amd.helpers import TARGET_TO_ARCH
 
 def custom_add_one(A:UOp) -> UOp:
   A = A.flatten()
@@ -43,9 +45,26 @@ def custom_add_var(A:UOp, B:UOp) -> UOp:
   sink = UOp.sink(A.base, B.base, var, threads, arg=KernelInfo(f"custom_add_var_{A.size}"))
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
 
+def custom_wave_sync(A:UOp, arch:str) -> UOp:
+  # 4 waves across 1024 WG — enough to saturate a SIMD with many concurrent WGs
+  # s_sleep yields the SIMD so waves from different WGs interleave, causing barrier packet reordering
+  threads = UOp.special(128, "lidx0")
+  wg = UOp.special(1024, "gidx0")
+  insts = []
+  for _ in range(4):
+    insts.append(s_sleep(4))
+    insts += [s_barrier()] if arch == "rdna3" else [s_barrier_signal(), s_barrier_wait()]
+    insts += [s_nop(0)]*4
+  insts.append(s_endpgm())
+  sink = UOp.sink(A.base, threads, wg, arg=KernelInfo("custom_wave_sync"))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+
 @unittest.skipUnless(Device.DEFAULT == "AMD", "requires AMD device")
 class TestCustomKernel(unittest.TestCase):
+  def setUp(self): self.arch = TARGET_TO_ARCH[Device["AMD"].arch]
+
   def test_simple(self):
+    if self.arch != "rdna3": self.skipTest("only rdna3")
     a = Tensor.full((16, 16), 1.).contiguous().realize()
     a = Tensor.custom_kernel(a, fxn=custom_add_one)[0]
     ei = a.schedule()[-1].lower()
@@ -55,6 +74,7 @@ class TestCustomKernel(unittest.TestCase):
     self.assertTrue((a.numpy() == 2.).all())
 
   def test_variable(self):
+    if self.arch != "rdna3": self.skipTest("only rdna3")
     b = Tensor.full((16, 16), 1, dtype=dtypes.uint32).contiguous().realize()
     a = Tensor.zeros_like(b).contiguous().realize()
     a = Tensor.custom_kernel(a, b, fxn=custom_add_var)[0]
@@ -63,5 +83,9 @@ class TestCustomKernel(unittest.TestCase):
       ei.run({"var":i})
       self.assertTrue((a.numpy() == 1+i).all())
 
+  def test_wave_sync(self):
+    if self.arch not in {"rdna3", "rdna4"}: self.skipTest("only rdna3 or rdna4")
+    Tensor.empty(1).custom_kernel(fxn=functools.partial(custom_wave_sync, arch=self.arch))[0].realize()
+
 if __name__ == "__main__":
   unittest.main()

test/amd/test_sqtt_examples.py

@@ -9,8 +9,8 @@ from tinygrad.renderer.amd import decode_inst
 from tinygrad.runtime.autogen.amd.rdna3.ins import SOPP
 from tinygrad.runtime.autogen.amd.rdna3.enum import SOPPOp
 from tinygrad.renderer.amd.sqtt import (decode, LAYOUT_HEADER, WAVESTART, WAVESTART_RDNA4, WAVEEND, INST, INST_RDNA4, VALUINST,
-                                     IMMEDIATE, IMMEDIATE_MASK, PACKET_TYPES_RDNA3, PACKET_TYPES_RDNA4,
-                                     InstOp, InstOpRDNA4, print_packets)
+                                     IMMEDIATE, IMMEDIATE_MASK, PACKET_TYPES_RDNA3, PACKET_TYPES_RDNA4, PACKET_TYPES_CDNA, CDNA_WAVESTART,
+                                     InstOp, InstOpRDNA4, print_packets, CDNA_WAVEEND, CDNA_INST)
 from test.amd.helpers import TARGET_TO_ARCH
 
 import tinygrad
@@ -21,7 +21,7 @@ OTHER_SIMD_OPS = {InstOp.OTHER_LDS_LOAD, InstOp.OTHER_LDS_STORE, InstOp.OTHER_LD
                   InstOp.OTHER_FLAT_STORE_128, InstOp.OTHER_GLOBAL_LOAD, InstOp.OTHER_GLOBAL_LOAD_VADDR,
                   InstOp.OTHER_GLOBAL_STORE_64, InstOp.OTHER_GLOBAL_STORE_96, InstOp.OTHER_GLOBAL_STORE_128,
                   InstOp.OTHER_GLOBAL_STORE_VADDR_128}
-OTHER_SIMD_OPS_RDNA4 = {InstOpRDNA4.OTHER_VMEM, InstOpRDNA4.OTHER_VMEM_5}
+OTHER_SIMD_OPS_RDNA4 = {InstOpRDNA4.OTHER_VMEM, InstOpRDNA4.OTHER_VMEM_5, InstOpRDNA4.OTHER_LDS_1, InstOpRDNA4.OTHER_LDS_2}
 
 # ═══════════════════════════════════════════════════════════════════════════════
 # ROCPROF DECODER
@@ -125,7 +125,7 @@ class SQTTExamplesTestBase(unittest.TestCase):
           self.assertIsInstance(packets[0], LAYOUT_HEADER, f"first packet should be LAYOUT_HEADER in {name}")
 
   def test_packet_types_valid(self):
-    all_classes = set(PACKET_TYPES_RDNA3.values()) | set(PACKET_TYPES_RDNA4.values())
+    all_classes = set(PACKET_TYPES_RDNA3.values()) | set(PACKET_TYPES_RDNA4.values()) | set(PACKET_TYPES_CDNA.values())
     for name, (events, *_) in self.examples.items():
       for i, event in enumerate(events):
         with self.subTest(example=name, event=i):
@@ -138,8 +138,8 @@ class SQTTExamplesTestBase(unittest.TestCase):
       if "empty" in name: continue
       with self.subTest(example=name):
         all_packets = [p for e in events for p in decode(e.blob)]
-        self.assertGreater(len([p for p in all_packets if isinstance(p, (WAVESTART, WAVESTART_RDNA4))]), 0, f"no WAVESTART in {name}")
-        self.assertGreater(len([p for p in all_packets if isinstance(p, WAVEEND)]), 0, f"no WAVEEND in {name}")
+        self.assertGreater(len([p for p in all_packets if isinstance(p, (WAVESTART, WAVESTART_RDNA4, CDNA_WAVESTART))]), 0, f"no WAVESTART in {name}")
+        self.assertGreater(len([p for p in all_packets if isinstance(p, (WAVEEND, CDNA_WAVEEND))]), 0, f"no WAVEEND in {name}")
 
   def test_time_monotonic(self):
     for name, (events, *_) in self.examples.items():
@@ -153,9 +153,10 @@ class SQTTExamplesTestBase(unittest.TestCase):
       if "gemm" not in name: continue
       with self.subTest(example=name):
         all_packets = [p for e in events for p in decode(e.blob)]
-        inst_names = [p.op.name for p in all_packets if isinstance(p, (INST, INST_RDNA4))]
-        self.assertGreater(len(inst_names), 0, f"no INST packets in {name}")
-        self.assertGreater(len([n for n in inst_names if n.startswith("JUMP")]), 0, f"no JUMP packets in {name}")
+        inst_packets = [p for p in all_packets if isinstance(p, (INST, INST_RDNA4, CDNA_INST))]
+        self.assertGreater(len(inst_packets), 0, f"no INST packets in {name}")
+        if isinstance(inst_packets[0], (INST, INST_RDNA4)):
+          self.assertGreater(len([p for p in inst_packets if p.op.name.startswith("JUMP")]), 0, f"no JUMP packets in {name}")
 
   expected: dict[str, list[int]] = {}  # override in subclasses
   def test_packet_counts(self):
@@ -183,8 +184,8 @@ class SQTTExamplesTestBase(unittest.TestCase):
         for event in events:
           wave_starts: dict[tuple[int, int, int], int] = {}
           for p in decode(event.blob):
-            if isinstance(p, (WAVESTART, WAVESTART_RDNA4)): wave_starts[(p.wave, p.simd, p.cu)] = p._time
-            elif isinstance(p, WAVEEND) and (key := (p.wave, p.simd, p.cu)) in wave_starts:
+            if isinstance(p, (WAVESTART, CDNA_WAVESTART, WAVESTART_RDNA4)): wave_starts[(p.wave, p.simd, p.cu)] = p._time
+            elif isinstance(p, (WAVEEND, CDNA_WAVEEND)) and (key := (p.wave, p.simd, p.cu)) in wave_starts:
               our_waves.append((wave_starts[key], p._time))
         self.assertEqual(sorted(our_waves), sorted(roc_waves), f"wave times mismatch in {name}")
 
@@ -210,19 +211,22 @@ class SQTTExamplesTestBase(unittest.TestCase):
 class TestSQTTExamplesRDNA3(SQTTExamplesTestBase):
   target = "gfx1100"
   expected = {
-    "profile_empty_run_0": [1974, 1961, 2014, 2065, 2092, 1998],
-    "profile_empty_run_1": [1979, 1972, 2019, 2070, 2097, 2003],
-    "profile_gemm_run_0": [2038, 11076, 2324, 2129, 2156, 2062],
-    "profile_gemm_run_1": [2038, 11037, 2318, 2129, 2156, 2062],
-    "profile_ops_run_0": [2038, 5070, 2078, 2129, 2156, 2062],
-    "profile_ops_run_1": [2038, 5007, 2078, 2129, 2156, 2062],
-    "profile_plus_run_0": [1979, 1979, 2030, 2070, 2097, 2003],
-    "profile_plus_run_1": [1979, 2043, 2030, 2070, 2097, 2003],
+    "profile_empty_run_0": [1880, 1867, 1920, 1971, 1998, 1904],
+    "profile_empty_run_1": [1880, 1867, 1920, 1971, 1998, 1904],
+    "profile_gemm_run_0": [3275, 3278, 2426, 2475, 2511, 2431],
+    "profile_gemm_run_1": [3264, 3268, 2420, 2469, 2504, 2401],
+    "profile_ops_run_0": [1944, 4903, 1984, 2035, 2062, 1968],
+    "profile_ops_run_1": [1944, 4918, 1984, 2035, 2062, 1968],
+    "profile_plus_run_0": [1938, 1932, 1978, 2029, 2056, 1962],
+    "profile_plus_run_1": [1891, 1874, 1931, 1982, 2009, 1915],
   }
 
 class TestSQTTExamplesRDNA4(SQTTExamplesTestBase): target = "gfx1200"
-@unittest.skip("TODO: fix CDNA")
-class TestSQTTExamplesCDNA(SQTTExamplesTestBase): target = "gfx950"
+
+class TestSQTTExamplesCDNA(SQTTExamplesTestBase):
+  target = "gfx950"
+  def test_rocprof_wave_times_match(self): self.skipTest("TODO: requires timestamp patching")
+  def test_rocprof_inst_times_match(self): self.skipTest("TODO: requires timestamp patching")
 
 if __name__ == "__main__":
   unittest.main()

test/amd/test_sqttmap.py

@@ -2,15 +2,16 @@
 import unittest, pickle
 from typing import Iterator
 from pathlib import Path
-from tinygrad.helpers import DEBUG, OSX
+from tinygrad.helpers import DEBUG, getenv, temp
 from tinygrad.renderer.amd.sqtt import print_packets, map_insts
 from tinygrad.runtime.autogen.amd.rdna3.ins import s_endpgm
+from tinygrad.viz.serve import sqtt_timeline
 from test.amd.disasm import disasm
 
 import tinygrad
 EXAMPLES_DIR = Path(tinygrad.__file__).parent.parent / "extra/sqtt/examples"
 
-def rocprof_inst_traces_match(sqtt, prg, target, pass_rocprof_err=False):
+def rocprof_inst_traces_match(sqtt, prg, target):
   from tinygrad.viz.serve import amd_decode
   from extra.sqtt.roc import decode as roc_decode, InstExec
   addr_table = amd_decode(prg.lib, target)
@@ -30,7 +31,7 @@ def rocprof_inst_traces_match(sqtt, prg, target, pass_rocprof_err=False):
     rocprof_inst = next(rwaves_iter[info.wave][0])
     ref_pc = rocprof_inst.pc-prg.base
     # always check pc matches
-    assert ref_pc == info.pc or pass_rocprof_err, f"pc mismatch {ref_pc}:{disasm_map[rocprof_inst.pc]} != {info.pc}:{disasm(info.inst)}"
+    assert ref_pc == info.pc, f"pc mismatch {ref_pc}:{disasm_map[rocprof_inst.pc]} != {info.pc}:{disasm(info.inst)}"
     # special handling for s_endpgm, it marks the wave completion.
     if info.inst == s_endpgm():
       completed_wave = list(rwaves_iter[info.wave].pop(0))
@@ -53,7 +54,7 @@ class TestSQTTMapBase(unittest.TestCase):
   def setUpClass(cls):
     if cls is TestSQTTMapBase: raise unittest.SkipTest("base class")
     cls.examples = {}
-    for pkl_path in sorted((EXAMPLES_DIR/cls.target).glob("*.pkl")):
+    for pkl_path in ([Path(temp("profile.pkl", append_user=True))] if getenv("LOAD_PROFILE") else sorted((EXAMPLES_DIR/cls.target).glob("*.pkl"))):
       with open(pkl_path, "rb") as f:
         data = pickle.load(f)
       sqtt_events = [e for e in data if type(e).__name__ == "ProfileSQTTEvent"]
@@ -63,18 +64,56 @@ class TestSQTTMapBase(unittest.TestCase):
 
   def test_rocprof_inst_traces_match(self):
     for name, (events, kern_events, target) in self.examples.items():
+      if "sync" in name and self.target.startswith("gfx12"):
+        self.skipTest("our timestamps are off by a few cycles because rocprof patches timestamps for rdna4 barriers")
       for event in events:
         if not event.itrace: continue
         if event.kern not in kern_events: continue
         with self.subTest(example=name, kern=event.kern):
-          # rocprof OSX has a bug for sopk decoding, linux rocprof works
-          pass_rocprof_err = OSX and target == "gfx1200" and name.startswith("profile_ops")
-          passed_insts, n_waves, n_units = rocprof_inst_traces_match(event, kern_events[event.kern], target, pass_rocprof_err)
+          passed_insts, n_waves, n_units = rocprof_inst_traces_match(event, kern_events[event.kern], target)
           if n_waves: print(f"{name}: passed for {passed_insts} instructions across {n_waves} waves scheduled on {n_units} wave units")
 
+  def test_sqtt_timeline(self):
+    for name, (events, kern_events, target) in self.examples.items():
+      for event in events:
+        if (p:=kern_events.get(event.kern)) is None: continue
+        with self.subTest(example=name, kern=event.kern):
+          # skip if there's no SQTT frequency data
+          if not (timeline:=sqtt_timeline(event.blob, p.lib, target)): continue
+          if not (frequency:=[e.key for e in timeline if type(e).__name__ == "ProfilePointEvent" and e.name == "freq_hz"]): continue
+          mean = sum(frequency) / len(frequency)
+          variance = sum((v - mean) ** 2 for v in frequency) / len(frequency)
+          self.assertGreater(mean, 0)
+          if DEBUG >= 2: print(f"{name:20s} SE:{event.se} {mean/1e9:.2f} GHz mean, {variance/1e18:.2f} GHz^2 variance")
+          events = [e for e in timeline if type(e).__name__ == "ProfileRangeEvent"]
+          insts, execs = 0, 0
+          for e in events:
+            if "EXEC" in e.device:
+              if "ALT" not in e.name.display_name: execs += 1
+            elif "WAVE" in e.device:
+              # sopk/immediates don't get ALU/MEM EXEC
+              if e.name.display_name not in {"IMMEDIATE", "IMMEDIATE_MASK", "JUMP", "JUMP_NO", "MESSAGE", "BARRIER", "BARRIER_SIGNAL"}: insts += 1
+            else: raise Exception(f"timeline row must be INST or EXEC, got {e.device}")
+          self.assertEqual(execs, insts)
+
+  def test_wave_sync(self):
+    for name, (events, kern_events, target) in self.examples.items():
+      for event in events:
+        wave_barriers = {}
+        for e in sqtt_timeline(event.blob, kern_events[event.kern].lib, target):
+          if type(e).__name__ == "ProfileRangeEvent" and e.name.display_name == "BARRIER": wave_barriers.setdefault(e.device, []).append(e)
+        if not wave_barriers: continue
+        for row, events in wave_barriers.items():
+          for e in events:
+            assert e.en-e.st > 1, f"all barriers must have a duration greater than 1, got {e}"
+
 class TestSQTTMapRDNA3(TestSQTTMapBase): target = "gfx1100"
 
 class TestSQTTMapRDNA4(TestSQTTMapBase): target = "gfx1200"
 
+class TestSQTTMapCDNA(TestSQTTMapBase):
+  target = "gfx950"
+  def test_rocprof_inst_traces_match(self): self.skipTest("requires timestamp patching to match rocprof, currently it's off by a few cycles")
+
 if __name__ == "__main__":
   unittest.main()

test/backend/test_arange.py

@@ -8,6 +8,7 @@ from tinygrad.engine.schedule import ExecItem
 from tinygrad.uop.ops import Ops
 from tinygrad.renderer import Estimates
 from tinygrad.renderer.ptx import PTXRenderer
+from test.helpers import needs_second_gpu
 
 class TestArange(unittest.TestCase):
   def _get_flops(self, tensor, desired):
@@ -188,6 +189,33 @@ 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)
 
+  @needs_second_gpu
+  @unittest.skipIf(Device.DEFAULT not in ("CPU", "AMD"), "atomics only on AMD/CPU")
+  @Context(USE_ATOMICS=1, SPEC=1)
+  def test_embedding_backward_vocab_sharded(self):
+    from tinygrad.renderer.cstyle import CStyleLanguage
+    if Device.DEFAULT == "CPU" and not isinstance(Device["CPU"].renderer, CStyleLanguage): self.skipTest("CPU needs Clang renderer")
+    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
+    vocab_size, embed_size = 1000, 128
+    bs, seqlen = 4, 256
+    idx = Tensor.randint(bs, seqlen, high=vocab_size)
+    emb = nn.Embedding(vocab_size, embed_size)
+    emb.weight = Tensor.ones(vocab_size, embed_size, requires_grad=True)
+    gt = Tensor.zeros(bs, seqlen, embed_size)
+    Tensor.realize(idx, emb.weight, gt)
+    # compute expected grad on single device
+    expected_grad = np.zeros((vocab_size, embed_size), dtype=np.float32)
+    for i in idx.flatten().numpy(): expected_grad[i] += 2
+    # now shard the embedding weight on vocab axis and recompute
+    emb.weight = Tensor.ones(vocab_size, embed_size, requires_grad=True)
+    emb.weight.shard_(devices, axis=0)
+    idx = idx.shard(devices, axis=None)
+    gt = gt.shard(devices, axis=None)
+    Tensor.realize(idx, emb.weight, gt)
+    loss = (emb(idx)-gt).square().sum()
+    loss.backward()
+    np.testing.assert_allclose(emb.weight.grad.numpy(), expected_grad, rtol=1e-5, atol=1e-5)
+
   @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):
     from extra.models.llama import precompute_freqs_cis, apply_rotary_emb

test/backend/test_asm_gemm.py

@@ -2,7 +2,7 @@ import unittest
 from tinygrad import Tensor, Device, dtypes, Context
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv
-from extra.gemm.asm.cdna.gemm import asm_gemm
+from extra.gemm.cdna_asm_gemm import asm_gemm
 from test.helpers import needs_second_gpu
 
 # On non CDNA4 it will only validate the Tensor.custom_kernel integration
@@ -87,6 +87,9 @@ class TestGemmLarge(unittest.TestCase):
     if not is_cdna4():
       self.skipTest("very slow on non mi350x")
 
+  @Context(ASM_GEMM=1)
+  def test_empty(self): (Tensor.empty(N:=getenv("N", 4096), N, dtype=dtypes.half)@Tensor.empty(N, N, dtype=dtypes.half)).realize()
+
   def test_tiny(self): verify_asm_gemm(1, 256, 256, 64)
   def test_simple(self): verify_asm_gemm(1, N:=getenv("N", 4096), N, N, dtype=dtypes.half)
   def test_gemm(self): verify_asm_gemm(1, 8192, 4096, 14336)
@@ -157,7 +160,7 @@ class TestGemmLarge(unittest.TestCase):
 
 class TestMagicGu(unittest.TestCase):
   def test_magicgu_matches_old(self):
-    from extra.gemm.asm.cdna.asm import _magicgu_mulhi, TILE_M, TILE_N, TILE_K
+    from extra.gemm.cdna_asm_gemm import _magicgu_mulhi, TILE_M, TILE_N, TILE_K
     old_iters_args = {64: (67108864, 0), 128: (33554432, 0), 224: (613566757, 2147483656)}
     old_gemm_shapes = [
       (8192, 4096, 4096), (8192, 14336, 4096), (8192, 4096, 14336),

test/backend/test_dtype.py

@@ -10,7 +10,7 @@ from tinygrad.renderer.nir import NIRRenderer
 from tinygrad import Context, Device, Tensor, dtypes
 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
+from test.unit.test_dtype_spec import _assert_eq, core_dtypes, dtype_ints, dtype_floats, FP8E4M3_MAX, FP8E5M2_MAX, FP8E4M3FNUZ_MAX, FP8E5M2FNUZ_MAX
 import pytest
 pytestmark = pytest.mark.filterwarnings("ignore")
 
@@ -101,14 +101,14 @@ class TestDType(unittest.TestCase):
   @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)), "skip for now")
   def test_uint_overflow(self):
     if not dtypes.is_unsigned(self.DTYPE): raise unittest.SkipTest("only for unsigned")
-    v = dtypes.max(self.DTYPE)
+    v = self.DTYPE.max
     _test_to_np(Tensor(v, dtype=self.DTYPE)+2, _to_np_dtype(self.DTYPE), np.array(v, dtype=_to_np_dtype(self.DTYPE))+2)
     _test_to_np(Tensor(v, dtype=self.DTYPE)*2, _to_np_dtype(self.DTYPE), np.array(v, dtype=_to_np_dtype(self.DTYPE))*2)
 
   def test_dtypes_DTYPES_DICT(self):
     self.assertIn("float", DTYPES_DICT)
     self.assertIn("float32", DTYPES_DICT)
-    self.assertEqual(len(DTYPES_DICT), 26)
+    self.assertEqual(len(DTYPES_DICT), 28)
     self.assertTrue(all(isinstance(value, DType) for value in DTYPES_DICT.values()))
     self.assertTrue(all(issubclass(_to_np_dtype(value), np.generic) for value in DTYPES_DICT.values() if _to_np_dtype(value) is not None))
 
@@ -143,6 +143,8 @@ def _test_ops(a_dtype:DType, b_dtype:DType, target_dtype=None):
 class TestFp8s(unittest.TestCase):
   def test_fp8e4m3_creation(self): assert Tensor([-1, 1, 2], dtype=dtypes.fp8e4m3).dtype == dtypes.fp8e4m3
   def test_fp8e5m2_creation(self): assert Tensor([-1, 1, 2], dtype=dtypes.fp8e5m2).dtype == dtypes.fp8e5m2
+  def test_fp8e4m3fnuz_creation(self): assert Tensor([-1, 1, 2], dtype=dtypes.fp8e4m3fnuz).dtype == dtypes.fp8e4m3fnuz
+  def test_fp8e5m2fnuz_creation(self): assert Tensor([-1, 1, 2], dtype=dtypes.fp8e5m2fnuz).dtype == dtypes.fp8e5m2fnuz
 
 class TestFp8sConversions(unittest.TestCase):
   @given(strat.floats(width=32, allow_subnormal=True, allow_nan=False, allow_infinity=False, min_value=-FP8E4M3_MAX, max_value=FP8E4M3_MAX))
@@ -169,6 +171,30 @@ class TestFp8sConversions(unittest.TestCase):
   def test_fp8e5m2_to_float(self, x):
     np.testing.assert_equal(fp8_to_float(x, dtypes.fp8e5m2), torch.tensor(x, dtype=torch.uint8).view(torch.float8_e5m2).float().item())
 
+  @given(strat.floats(width=32, allow_subnormal=True, allow_nan=False, allow_infinity=False, min_value=-FP8E4M3FNUZ_MAX, max_value=FP8E4M3FNUZ_MAX))
+  def test_float_to_fp8e4m3fnuz(self, x):
+    np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e4m3fnuz), torch.tensor(x, dtype=torch.float8_e4m3fnuz).view(torch.uint8).item())
+
+  def test_float_to_fp8e4m3fnuz_extreme_values(self):
+    for x in [FP8E4M3FNUZ_MAX, FP8E4M3FNUZ_MAX*1.01, -FP8E4M3FNUZ_MAX, -FP8E4M3FNUZ_MAX*1.01, math.inf, -math.inf, math.nan, 0.0, -0.0]:
+      np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e4m3fnuz), torch.tensor(x, dtype=torch.float8_e4m3fnuz).view(torch.uint8).item())
+
+  @given(strat.floats(width=32, allow_subnormal=True, allow_nan=False, allow_infinity=False, min_value=-FP8E5M2FNUZ_MAX, max_value=FP8E5M2FNUZ_MAX))
+  def test_float_to_fp8e5m2fnuz(self, x):
+    np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e5m2fnuz), torch.tensor(x, dtype=torch.float8_e5m2fnuz).view(torch.uint8).item())
+
+  def test_float_to_fp8e5m2fnuz_extreme_values(self):
+    for x in [FP8E5M2FNUZ_MAX, FP8E5M2FNUZ_MAX*1.01, -FP8E5M2FNUZ_MAX, -FP8E5M2FNUZ_MAX*1.01, math.inf, -math.inf, math.nan, 0.0, -0.0]:
+      np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e5m2fnuz), torch.tensor(x, dtype=torch.float8_e5m2fnuz).view(torch.uint8).item())
+
+  @given(strat.integers(min_value=0, max_value=255))
+  def test_fp8e4m3fnuz_to_float(self, x):
+    np.testing.assert_equal(fp8_to_float(x, dtypes.fp8e4m3fnuz), torch.tensor(x, dtype=torch.uint8).view(torch.float8_e4m3fnuz).float().item())
+
+  @given(strat.integers(min_value=0, max_value=255))
+  def test_fp8e5m2fnuz_to_float(self, x):
+    np.testing.assert_equal(fp8_to_float(x, dtypes.fp8e5m2fnuz), torch.tensor(x, dtype=torch.uint8).view(torch.float8_e5m2fnuz).float().item())
+
 @unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), "bfloat16 not supported")
 class TestBFloat16(unittest.TestCase):
   def test_bf16_creation_numpy(self):
@@ -490,4 +516,3 @@ class TestOpsBFloat16(unittest.TestCase):
 
 if __name__ == '__main__':
   unittest.main()
-

test/backend/test_dtype_alu.py

@@ -52,6 +52,8 @@ class ht:
 ht.bfloat16 = ht.uint16.filter(lambda x: ((x >> 7) & 0xFF) != 0)  # filter subnormal bfloat16
 ht.fp8e4m3 = ht.uint8
 ht.fp8e5m2 = ht.uint8
+ht.fp8e4m3fnuz = ht.uint8
+ht.fp8e5m2fnuz = ht.uint8
 
 def universal_test(a, b, dtype, op):
   if not isinstance(op, tuple): op = (op, op)
@@ -67,7 +69,8 @@ def universal_test(a, b, dtype, op):
     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))
+    else: atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1),
+                        dtypes.fp8e4m3fnuz:(1e-1, 1e-1), dtypes.fp8e5m2fnuz:(5e-1, 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)
 
@@ -87,7 +90,8 @@ def universal_test_unary(a, dtype, op):
   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))
+      dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2: (1.0, 5e-1),
+      dtypes.fp8e4m3fnuz:(1e-1, 1e-1), dtypes.fp8e5m2fnuz: (5e-1, 5e-1)}.get(dtype, (1e-6, 1e-5))
     np.testing.assert_allclose(tensor_value, numpy_value, atol=atol, rtol=rtol)
   else: np.testing.assert_equal(tensor_value, numpy_value)
 
@@ -155,6 +159,26 @@ class TestDTypeALU(unittest.TestCase):
   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)
 
+  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3fnuz), f"no fp8e4m3fnuz on {Device.DEFAULT}")
+  @given(ht.fp8e4m3fnuz, ht.fp8e4m3fnuz, strat.sampled_from(binary_operations))
+  def test_fp8e4m3fnuz(self, a, b, op):
+    universal_test(from_storage_scalar(a, dtypes.fp8e4m3fnuz), from_storage_scalar(b, dtypes.fp8e4m3fnuz), dtypes.fp8e4m3fnuz, op)
+
+  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e5m2fnuz), f"no fp8e5m2fnuz on {Device.DEFAULT}")
+  @given(ht.fp8e5m2fnuz, ht.fp8e5m2fnuz, strat.sampled_from(binary_operations))
+  def test_fp8e5m2fnuz(self, a, b, op):
+    universal_test(from_storage_scalar(a, dtypes.fp8e5m2fnuz), from_storage_scalar(b, dtypes.fp8e5m2fnuz), dtypes.fp8e5m2fnuz, op)
+
+  @given(ht.fp8e4m3fnuz, ht.fp8e4m3fnuz, strat.sampled_from(binary_operations))
+  @Context(EMULATED_DTYPES="fp8e4m3fnuz")
+  def test_emulated_fp8e4m3fnuz(self, a, b, op):
+    universal_test(from_storage_scalar(a, dtypes.fp8e4m3fnuz), from_storage_scalar(b, dtypes.fp8e4m3fnuz), dtypes.fp8e4m3fnuz, op)
+
+  @given(ht.fp8e5m2fnuz, ht.fp8e5m2fnuz, strat.sampled_from(binary_operations))
+  @Context(EMULATED_DTYPES="fp8e5m2fnuz")
+  def test_emulated_fp8e5m2fnuz(self, a, b, op):
+    universal_test(from_storage_scalar(a, dtypes.fp8e5m2fnuz), from_storage_scalar(b, dtypes.fp8e5m2fnuz), dtypes.fp8e5m2fnuz, op)
+
   @given(ht.float32, strat.sampled_from(unary_operations))
   def test_float32_unary(self, a, op): universal_test_unary(a, dtypes.float32, op)
 
@@ -198,6 +222,30 @@ class TestDTypeALU(unittest.TestCase):
     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)
 
+  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3fnuz), f"no fp8e4m3fnuz on {Device.DEFAULT}")
+  @given(ht.fp8e4m3fnuz, strat.sampled_from(unary_operations))
+  def test_fp8e4m3fnuz_unary(self, a, op):
+    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e4m3fnuz) != 0.0)
+    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e4m3fnuz), dtypes.fp8e4m3fnuz, op)
+
+  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e5m2fnuz), f"no fp8e5m2fnuz on {Device.DEFAULT}")
+  @given(ht.fp8e5m2fnuz, strat.sampled_from(unary_operations))
+  def test_fp8e5m2fnuz_unary(self, a, op):
+    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e5m2fnuz) != 0.0)
+    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e5m2fnuz), dtypes.fp8e5m2fnuz, op)
+
+  @given(ht.fp8e4m3fnuz, strat.sampled_from(unary_operations))
+  @Context(EMULATED_DTYPES="fp8e4m3fnuz")
+  def test_emulated_fp8e4m3fnuz_unary(self, a, op):
+    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e4m3fnuz) != 0.0)
+    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e4m3fnuz), dtypes.fp8e4m3fnuz, op)
+
+  @given(ht.fp8e5m2fnuz, strat.sampled_from(unary_operations))
+  @Context(EMULATED_DTYPES="fp8e5m2fnuz")
+  def test_emulated_fp8e5m2fnuz_unary(self, a, op):
+    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e5m2fnuz) != 0.0)
+    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e5m2fnuz), dtypes.fp8e5m2fnuz, 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)
 
@@ -318,7 +366,7 @@ class TestDTypeALU(unittest.TestCase):
 
   @unittest.expectedFailure
   def test_unsafe_cast_float_to_int_failure(self):
-    val = float(dtypes.max(dtypes.int32) - 1)
+    val = float(dtypes.int32.max - 1)
     t1 = Tensor([val], dtype=dtypes.float32).cast(dtypes.int32)
     t2 = Tensor(val, dtype=dtypes.float32).cast(dtypes.int32)
     np.testing.assert_equal(t1.item(), t2.item())

test/backend/test_ops.py

@@ -6,6 +6,7 @@ from tinygrad.helpers import getenv, IMAGE, DEBUG, CI, Context, CPU_LLVM, AMD_LL
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.device import is_dtype_supported
+from tinygrad.renderer.cstyle import QCOMCLRenderer
 from tinygrad.renderer.nir import NIRRenderer
 
 TINY_BACKEND = getenv("TINY_BACKEND")
@@ -436,7 +437,7 @@ class TestOps(unittest.TestCase):
     helper_test_op([(45,35), (45,35), (45,35)], lambda x,y,z: x.lerp(y,z))
     helper_test_op(None, lambda x,y,z: x.lerp(y,z), vals=[[1.,2.,3.], [4.,5.,6.], 0.5])
 
-  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_tril(self):
     helper_test_op([(3,3)], lambda x: x.tril())
     helper_test_op([(3,3)], lambda x: x.tril(1))
@@ -454,7 +455,7 @@ class TestOps(unittest.TestCase):
     helper_test_op([(5,3,3)], lambda x: x.tril(1))
     helper_test_op(None, lambda x: x.tril(), vals=[[[True] * 3] * 3], forward_only=True)
 
-  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_triu(self):
     helper_test_op([(3,3)], lambda x: x.triu())
     helper_test_op([(3,3)], lambda x: x.triu(1))
@@ -478,9 +479,9 @@ class TestOps(unittest.TestCase):
     helper_test_op(None, torch.maximum, Tensor.maximum, vals=[[1., 0., 3., -4.], 3.])
     helper_test_op(None, torch.maximum, Tensor.maximum, vals=[[1., 0., 3., -4.], [-1., -2., 3., 0.]])
     helper_test_op(None, torch.maximum, Tensor.maximum,
-                   vals=[[-1234, 0, 1234, dtypes.max(dtypes.int), dtypes.min(dtypes.int)], dtypes.max(dtypes.int)], forward_only=True)
+                   vals=[[-1234, 0, 1234, dtypes.int.max, dtypes.int.min], dtypes.int.max], forward_only=True)
     helper_test_op(None, torch.maximum, Tensor.maximum,
-                   vals=[[-1234, 0, 1234, dtypes.max(dtypes.int), dtypes.min(dtypes.int)], dtypes.min(dtypes.int)], forward_only=True)
+                   vals=[[-1234, 0, 1234, dtypes.int.max, dtypes.int.min], dtypes.int.min], forward_only=True)
     helper_test_op(None, torch.maximum, Tensor.maximum, vals=[[True, False, False], True], forward_only=True)
     helper_test_op(None, torch.maximum, Tensor.maximum, vals=[[True, False, False], [True, True, False]], forward_only=True)
 
@@ -495,9 +496,9 @@ class TestOps(unittest.TestCase):
     helper_test_op(None, torch.minimum, Tensor.minimum, vals=[[1., 0., 3., -4.], 3.])
     helper_test_op(None, torch.minimum, Tensor.minimum, vals=[[1., 0., 3., -4.], [-1., -2., 3., 0.]])
     helper_test_op(None, torch.minimum, Tensor.minimum,
-                   vals=[[-1234, 0, 1234, dtypes.max(dtypes.int), dtypes.min(dtypes.int)], dtypes.max(dtypes.int)], forward_only=True)
+                   vals=[[-1234, 0, 1234, dtypes.int.max, dtypes.int.min], dtypes.int.max], forward_only=True)
     helper_test_op(None, torch.minimum, Tensor.minimum,
-                   vals=[[-1234, 0, 1234, dtypes.max(dtypes.int), dtypes.min(dtypes.int)], dtypes.min(dtypes.int)], forward_only=True)
+                   vals=[[-1234, 0, 1234, dtypes.int.max, dtypes.int.min], dtypes.int.min], forward_only=True)
     helper_test_op(None, torch.minimum, Tensor.minimum, vals=[[True, False, False], True], forward_only=True)
     helper_test_op(None, torch.minimum, Tensor.minimum, vals=[[True, False, False], [True, True, False]], forward_only=True)
 
@@ -765,6 +766,7 @@ class TestOps(unittest.TestCase):
 
     self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_xor(y), expected=RuntimeError)
 
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_and(self):
     data = [[1,-8,1],[32,1,6]]
     tor = torch.tensor(data, dtype=torch.int)
@@ -782,6 +784,7 @@ class TestOps(unittest.TestCase):
 
     self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_and(y), expected=RuntimeError)
 
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_or(self):
     data = [[1,-8,1],[32,1,6]]
     tor = torch.tensor(data, dtype=torch.int)
@@ -1170,6 +1173,7 @@ class TestOps(unittest.TestCase):
     helper_test_op(None, lambda x: x.type(torch.int32).argmax().type(torch.int32), lambda x: x.argmax(), forward_only=True, vals=[[False, True]])
     helper_test_op(None, lambda x: x.type(torch.int32).argmax().type(torch.int32), lambda x: x.argmax(), forward_only=True, vals=[[True, False]])
 
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_argmin(self):
     # check if it returns the first index for multiple occurrences
     helper_test_op(None, lambda x: x.argmin().type(torch.int32), lambda x: x.argmin(), forward_only=True, vals=[[2, 2]])
@@ -1475,6 +1479,7 @@ class TestOps(unittest.TestCase):
   def test_prod_dtype_arg(self):
     with self.assertRaises(AttributeError): Tensor([1.0, 2.0]).prod(dtype="")
 
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_min(self):
     helper_test_op([(3,3)], lambda x: x.min())
     helper_test_op([(45,3)], lambda x: x.min())
@@ -1503,7 +1508,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(3,3)], lambda x: torch.full_like(x, 2).prod(), lambda x: (x.full_like(2)).prod(), forward_only=True)
     helper_test_op([(3,3)], lambda x: torch.full_like(x, 2).max(), lambda x: (x.full_like(2)).max(), forward_only=True)
 
-  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
   def test_any(self):
     helper_test_op([(3,4,5,6)], lambda x: x.any(), forward_only=True)
     helper_test_op(None, lambda x: x.any(), vals=[[True, True]], forward_only=True)
@@ -1515,7 +1519,7 @@ class TestOps(unittest.TestCase):
   def test_any_zero_axis(self):
     helper_test_op([(1,0,3,0,5)], lambda x: x.any(axis=(1,3)), forward_only=True)
 
-  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_all(self):
     helper_test_op([(3,4,5,6)], lambda x: x.all(), forward_only=True)
     helper_test_op(None, lambda x: x.all(), vals=[[True, True]], forward_only=True)
@@ -2889,6 +2893,7 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[...,c,:,e], lambda x: x[...,k,:,p])
 
   @slow_test
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_dim_collapse_int(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     # dim collapse from int
@@ -2899,6 +2904,7 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[1,:,3:11:2,d,0:2], lambda x: x[1,:,3:11:2,o,0:2])
 
   @slow_test
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_dim_inject_none(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     # dim injection from None
@@ -2933,6 +2939,7 @@ class TestOps(unittest.TestCase):
                             lambda x: x[Tensor([[0,1,-1],[-1,-2,0]]), Tensor([2,1,-1])])
 
   @slow_test
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_list_indices(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[((0,),)])
@@ -2944,6 +2951,7 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,(2,1,0),c,(-2,1,0),e], lambda x: x[i,(2,1,0),k,(-2,1,0),p])
 
   @slow_test
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_tuple_indices(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[(((0,),),)], lambda x: x[(((0,),),)])
@@ -3285,7 +3293,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(20,)], lambda x: (x>0.5).nonzero().int(), lambda x: (x>0.5).nonzero(), forward_only=True)
     helper_test_op([(10, 5, 3)], lambda x: (x>0.5).nonzero().int(), lambda x: (x>0.5).nonzero(), forward_only=True)
 
-  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
   def test_cast(self):
     helper_test_op([(3, 3)], lambda x: x.float())
     helper_test_op(None, lambda x: x.float(), vals=[[0, 1, 2, 3]], forward_only=True)

test/backend/test_quantize_onnx.py

@@ -204,13 +204,13 @@ class TestQuantizeOnnx(unittest.TestCase):
     W = Tensor(m2:=(np.random.uniform(0, 255, size=(N,N)).astype(wi))).realize()
     tg_dtype = dtypes.int8 if xi == np.int8 else dtypes.uint8
     out = (X.int().matmul(W.int())//1000)
-    if clip: out = out.clip(dtypes.min(tg_dtype),dtypes.max(tg_dtype))
+    if clip: out = out.clip(tg_dtype.min, tg_dtype.max)
     out = out.cast(tg_dtype)
     opts = [Opt(op=OptOps.UPCAST, axis=1, arg=128), Opt(op=OptOps.UNROLL, axis=0, arg=4)] if opts is None else opts
     sexec(out, opts, replace_src, run_count=1)
     tout = out.numpy()
     mout = ((m1.astype(np.int32) @ m2.astype(np.int32)) // 1000)
-    if clip: mout = mout.clip(dtypes.min(tg_dtype),dtypes.max(tg_dtype))
+    if clip: mout = mout.clip(tg_dtype.min, tg_dtype.max)
     mout = mout.astype(xi)
     print(tout)
     print(mout)

test/backend/test_renderer_failures.py

@@ -62,7 +62,7 @@ class TestRendererFailures(unittest.TestCase):
 class TestCStyleFailures(unittest.TestCase):
   def test_inline_const_alu(self):
     # CPU doesn't use the max function
-    ret = _setup_and_test_alu(Ops.MAX, 1, UOp.const(dtypes.int, dtypes.min(dtypes.int)+1))
+    ret = _setup_and_test_alu(Ops.MAX, 1, UOp.const(dtypes.int, dtypes.int.min+1))
     self.assertEqual(ret[0], 1)
 
   def _test_src_strip_paren(self, op: Ops, should_strip_paren:bool=True):

test/backend/test_schedule.py

@@ -811,6 +811,26 @@ class TestSchedule(unittest.TestCase):
       self.assertEqual(cnt1, 5)
       self.assertEqual(cnt2, 5)
 
+  @unittest.skipIf(Device.DEFAULT != "CL", "image only supported on CL")
+  def test_image_f16_residual_fusion(self):
+    with Context(FLOAT16=1, OPENPILOT_HACKS=1):
+      def cnt():
+        inp = Tensor.empty((512,), dtype='float')
+        b1, b2 = Tensor.empty((512, 1024), dtype='float'), Tensor.empty((1024, 512), dtype='float')
+        c1, c2 = Tensor.empty((1024,), dtype='float'), Tensor.empty((512,), dtype='float')
+        rb = (((((inp @ b1) + c1).relu() @ b2) + c2).relu() + inp).relu()
+        b16, c16 = Tensor.empty((512, 16), dtype='float'), Tensor.empty((16,), dtype='float')
+        b32, c32 = Tensor.empty((512, 32), dtype='float'), Tensor.empty((32,), dtype='float')
+        sched = Tensor.schedule((rb @ b16 + c16).relu(), (rb @ b32 + c32).relu())
+        for si in sched: si.lower()
+        return len([si for si in sched if isinstance(si.prg, CompiledRunner)])
+
+      with Context(IMAGE=1): cnt1 = cnt()
+      with Context(IMAGE=2): cnt2 = cnt()
+
+      self.assertEqual(cnt1, 9)
+      self.assertEqual(cnt2, 9)
+
   @unittest.skipIf(Device.DEFAULT != "CL", "image only supported on CL")
   @unittest.expectedFailure
   def test_image_conv_fusion(self):

test/mockgpu/amd/amdgpu.py

@@ -271,6 +271,7 @@ class SDMAExecutor(AMDQueue):
       elif op == amd_gpu.SDMA_OP_GCR: self._execute_gcr()
       elif op == amd_gpu.SDMA_OP_COPY: self._execute_copy()
       elif op == amd_gpu.SDMA_OP_TIMESTAMP: self._execute_timestamp()
+      elif op == 32: self.rptr[0] += 4  # SDMA_OP_DUMMY_TRAP: pipeline flush, no interrupt
       else: raise RuntimeError(f"Unknown SDMA op {op}")
     return self.rptr[0] - prev_rptr
 

test/null/test_dtype_spec.py

@@ -75,20 +75,20 @@ class TestHelpers(unittest.TestCase):
   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)
         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)
         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)
         np.testing.assert_equal(dt.min, False)
         np.testing.assert_equal(dt.max, True)
 

test/null/test_memory_planner.py

@@ -11,8 +11,8 @@ def b(i, base=None, offset=0, pin=False, size=16):
   if pin: global_map[i].ref(1)
   return global_map[i]
 
-def check_assign(buffers:list[list[Buffer]|tuple[Buffer, ...]]):
-  assigned = _internal_memory_planner(buffers, noopt_buffers=None)
+def check_assign(buffers:list[list[Buffer]|tuple[Buffer, ...]], copies:list[tuple[Buffer, Buffer]]|None=None):
+  assigned = _internal_memory_planner(buffers, copies=copies)
 
   taken_parts = set()
   first_appearance, last_appearance = {}, {}
@@ -134,5 +134,75 @@ class TestMemoryPlanner(unittest.TestCase):
     ]
     check_assign(bs)
 
+  def test_copy_bufs_separate_from_compute(self):
+    bs = [
+      [b(0), b(1)],
+      [b(1), b(2)],
+      [b(3), b(2)],
+    ]
+    assigned = _internal_memory_planner(bs, copies=[(b(1), b(0))])
+    r1, r2 = assigned.get(b(1), b(1)), assigned.get(b(2), b(2))
+    assert r1.base != r2.base
+
+  def test_copy_bufs_reuse_among_copies(self):
+    bs = [
+      [b(0), b(1)],
+      [b(2), b(1)],
+      [b(3), b(2)],
+    ]
+    assigned = _internal_memory_planner(bs, copies=[(b(1), b(0)), (b(2), b(1))])
+    r1, r2 = assigned.get(b(1), b(1)), assigned.get(b(2), b(2))
+    assert r1.base == r2.base
+
+  def test_compute_bufs_reuse_among_compute(self):
+    bs = [
+      [b(0), b(1)],
+      [b(2), b(1)],
+      [b(3), b(2)],
+      [b(4), b(3)],
+    ]
+    assigned = _internal_memory_planner(bs, copies=[(b(1), b(0))])
+    r2, r3 = assigned.get(b(2), b(2)), assigned.get(b(3), b(3))
+    assert r2.base == r3.base
+
+  def test_copy_and_compute_no_cross_reuse(self):
+    bs = [
+      [b(0), b(1)],
+      [b(2), b(1)],
+      [b(3), b(2)],
+    ]
+    assigned = _internal_memory_planner(bs, copies=[(b(2), b(1))])
+    r0, r2 = assigned.get(b(0), b(0)), assigned.get(b(2), b(2))
+    assert r0.base != r2.base
+
+  def test_multiple_copy_bufs_with_offsets(self):
+    bs = [
+      [b(0, pin=True), b(1), b(2)],
+      [b(3, base=0, offset=1, size=8), b(1), b(2)],
+      [b(4), b(3)],
+      [b(5), b(4)],
+    ]
+    check_assign(bs, copies=[(b(1), b(0)), (b(2), b(0))])
+
+  def test_copy_bufs_pinned_mixed(self):
+    bs = [
+      [b(0, pin=True), b(1), b(2)],
+      [b(1), b(3), b(2)],
+      [b(4), b(3)],
+      [b(5), b(4), b(0)],
+    ]
+    check_assign(bs, copies=[(b(1), b(0)), (b(3), b(1))])
+
+  def test_deferred_copy_frees_chain(self):
+    bs = []
+    copies = []
+    for i in range(6):
+      copy_buf, compute_buf = b(i * 2 + 1), b(i * 2 + 2)
+      bs.append([copy_buf, b(0, pin=True)])
+      bs.append([compute_buf, copy_buf])
+      copies.append((copy_buf, b(0, pin=True)))
+    bs.append([b(100, pin=True)])
+    check_assign(bs, copies=copies)
+
 if __name__ == "__main__":
   unittest.main()

test/null/test_multitensor.py

@@ -108,7 +108,6 @@ class TestMultiRamUsage(unittest.TestCase):
   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.expectedFailure
   def test_multi_layer_allreduce(self):
     N = 32
     devices_2 = ("NULL:1", "NULL:2")

test/null/test_schedule.py

@@ -1189,5 +1189,11 @@ class TestBufferView(unittest.TestCase):
     b = a.shrink(((200, 800),)).shrink(((0, 300),)).reshape((30, 10)).shrink(((20, 25), (0, 10))).contiguous()
     run_schedule(check_schedule(b, 0))
 
+class TestInvalidTensor(unittest.TestCase):
+  def test_full_invalid_is_zero_kernels(self):
+    from tinygrad.dtype import Invalid
+    t = Tensor.full((4,), Invalid, dtype=dtypes.float)
+    check_schedule(t, 0)
+
 if __name__ == '__main__':
   unittest.main(verbosity=2)

test/null/test_simplify_valid_idx.py

@@ -46,9 +46,9 @@ class TestHelpers(unittest.TestCase):
     self.assertTrue((rng+2).is_increasing())
 
 class TestValidIdxSimplification(unittest.TestCase):
-  def check(self, load, sidx, svalid):
+  def check(self, load, sidx, svalid, extra=()):
     with Context(NOOPT=1, SPEC=0):
-      load = full_rewrite_to_sink(load.sink()).src[0]
+      load = full_rewrite_to_sink(UOp.sink(load, *extra)).src[0]
     idx, valid = load.src[0].src[1], load.src[0].src[2]
     check_uop_against_string(self, idx, sidx)
     check_uop_against_string(self, valid, svalid)
@@ -156,9 +156,12 @@ class TestValidIdxSimplification(unittest.TestCase):
     idx = (alu15*-31)+(((((alu11+218)//224)+ridx0)%30)*1568)
     valid = (ridx2<1)&(ridx1<6)
     load = get_gated_load_uop(valid, idx)
+    # prevent ridx1 and ridx2 from being shrunk
+    red = UOp(Ops.REDUCE, dtypes.float, (load, ridx1, ridx2), Ops.ADD)
     self.check(load,
       "(r0*1568)",
-      "((r2<1)&(r1<6))")
+      "((r2<1)&(r1<6))",
+      extra=(red,))
 
   def test_valid_becomes_const1_z3(self):
     from z3 import Ints, Solver, And, If, Not, unsat
@@ -351,7 +354,7 @@ class TestImageSimplification(unittest.TestCase):
 
     self.check(load,
                "((((idx2*2)+r0)<11)&((((idx1*8)+r1)<3)!=True))",
-               "(((idx0+((idx1*512)+(r1*64)))+832)%1024)",
+               "(idx0+(idx1*512+r1*64)+-192)",
                "((((idx2*2)+r0)+(((idx1+((r1+5)//8))+1)//2))+-4)")
 
   def test_simplify1(self):
@@ -388,18 +391,17 @@ class TestImageSimplification(unittest.TestCase):
     alu8 = (idx0//8%32//4)
     alu9 = idx0<256
 
-    # TODO: can this be simplified further?
     load = get_load_image_uop(shape, alu9, (((alu8+(alu2*8))%64),(alu2//8)))
-    self.check(load, "(idx0<256)", "(((((idx0%8)*32)+(idx0//32))+8)%64)", "(idx0//2%4)")
+    self.check(load, "(idx0<256)", "(idx0%2*32+idx0//32+8)", "(idx0//2%4)")
 
     load = get_load_image_uop(shape, alu9, (((alu8+(alu3*8))%64),(alu3//8)))
-    self.check(load, "(idx0<256)", "(((((idx0%8)*32)+(idx0//32))+16)%64)", "(idx0//2%4)")
+    self.check(load, "(idx0<256)", "(idx0%2*32+idx0//32+16)", "(idx0//2%4)")
 
     load = get_load_image_uop(shape, alu9, (((alu8+(alu4*8))%64),(alu4//8)))
-    self.check(load, "(idx0<256)", "(((((idx0%8)*32)+(idx0//32))+24)%64)", "(idx0//2%4)")
+    self.check(load, "(idx0<256)", "(idx0%2*32+idx0//32+24)", "(idx0//2%4)")
 
     load = get_load_image_uop(shape, alu9, (((alu8+(alu5*8))%64),(alu5//8)))
-    self.check(load, "(idx0<256)", "((((idx0%8)*32)+(idx0//32))%64)", "(idx0//2%4)")
+    self.check(load, "(idx0<256)", "(idx0%2*32+idx0//32)", "(idx0//2%4)")
 
   def test_simplify5(self):
     # openpilot 0.9.7, chunk replacement to simplify
@@ -414,7 +416,7 @@ class TestImageSimplification(unittest.TestCase):
     valid = alu3<640
 
     load = get_load_image_uop(shape, valid, idx)
-    self.check(load, "(((idx0+(idx1*64))%192)<160)", "((idx0+((idx1//3)*16))+128)", "((idx1%3)*4)")
+    self.check(load, None, "((idx0+((idx1//3)*16))+128)", "((idx1%3)*4)")
 
   def test_simplify6(self):
     # from openpilot
@@ -484,5 +486,79 @@ class TestDropTrueGate(unittest.TestCase):
     # the True gate should be dropped (INDEX should only have 2 sources)
     self.assertEqual(len(result.src), 2, "True gate should be dropped from INDEX")
 
+class TestRangeShrink(unittest.TestCase):
+  def get_ranges(self, sink):
+    with Context(NOOPT=1, SPEC=0):
+      result = full_rewrite_to_sink(sink)
+    return [u for u in result.toposort() if u.op is Ops.RANGE]
+
+  def test_range_shrink_single_guard(self):
+    # range 0..203 guarded by r < 4 everywhere -> shrink to 0..3
+    r = Range(0, 204)
+    load = get_gated_load_uop(r < UOp.const(dtypes.index, 4), r)
+    ranges = self.get_ranges(load.sink())
+    self.assertEqual(len(ranges), 1)
+    self.assertEqual(ranges[0].src[0].arg, 4)
+
+  def test_range_shrink_picks_max_guard(self):
+    # two loads guard the same range with r < 4 and r < 8 -> shrink to max(4, 8) = 8
+    r = Range(0, 204)
+    load1 = get_gated_load_uop(r < UOp.const(dtypes.index, 4), r)
+    load2 = get_gated_load_uop(r < UOp.const(dtypes.index, 8), r)
+    ranges = self.get_ranges(UOp.sink(load1, load2))
+    self.assertEqual(len(ranges), 1)
+    self.assertEqual(ranges[0].src[0].arg, 8)
+
+  def test_range_no_shrink_guard_ge_max(self):
+    # guard r < 300 with range max 204 -> no shrink (guard doesn't constrain)
+    r = Range(0, 204)
+    load = get_gated_load_uop(r < UOp.const(dtypes.index, 300), r)
+    ranges = self.get_ranges(load.sink())
+    self.assertEqual(len(ranges), 1)
+    self.assertEqual(ranges[0].src[0].arg, 204)
+
+  def test_range_no_shrink_when_unguarded_elsewhere(self):
+    # one load guards r < 4, but another load uses r without a gate -> no shrink
+    r = Range(0, 204)
+    load1 = get_gated_load_uop(r < UOp.const(dtypes.index, 4), r)
+    load2 = UOp(Ops.LOAD, dtypes.float, (UOp(Ops.PARAM, dtypes.float.ptr(), arg=1).index(r, ptr=True),))
+    ranges = self.get_ranges(UOp.sink(load1, load2))
+    self.assertEqual(len(ranges), 1)
+    self.assertEqual(ranges[0].src[0].arg, 204)
+
+  def test_range_no_shrink_when_used_in_reduce(self):
+    # range used in both a gated load AND directly in the reduce expression -> no shrink
+    r = Range(0, 204)
+    gated_load = get_gated_load_uop(r < UOp.const(dtypes.index, 4), r)
+    red = UOp(Ops.REDUCE, dtypes.float, (r.cast(dtypes.float) + gated_load, r), Ops.ADD)
+    ranges = self.get_ranges(red.sink())
+    self.assertEqual(len(ranges), 1)
+    self.assertEqual(ranges[0].src[0].arg, 204)
+
+  def test_range_shrink_to_single_iteration(self):
+    # guard r < 1 shrinks range to 1 -> single iteration, range eliminated entirely
+    r = Range(0, 204)
+    load = get_gated_load_uop(r < UOp.const(dtypes.index, 1), r)
+    ranges = self.get_ranges(load.sink())
+    self.assertEqual(len(ranges), 0)
+
+  def test_range_shrink_store_where_invalid(self):
+    # emulates mask.where(x.pad_to(mask.shape), Invalid): range should shrink accordingly
+    from tinygrad.dtype import Invalid
+    r = Range(0, 204)
+    x = (r < 4).where(UOp.const(dtypes.float, 1), Invalid)
+    ranges = self.get_ranges(UOp(Ops.PARAM, dtypes.float.ptr(), arg=0).index(r).store((r < 4).where(x, 0)).sink())
+    self.assertEqual(len(ranges), 1)
+    self.assertEqual(ranges[0].src[0].arg, 4)
+
+  def test_range_shrink_store_where_invalid_flipped(self):
+    # above, but flipped
+    from tinygrad.dtype import Invalid
+    r = Range(0, 204)
+    x = (r < 4).where(UOp.const(dtypes.float, 1), Invalid)
+    ranges = self.get_ranges(UOp(Ops.PARAM, dtypes.float.ptr(), arg=0).index(r).store((r < 4).where(0, x)).sink())
+    self.assertEqual(len(ranges), 1)
+    self.assertEqual(ranges[0].src[0].arg, 4)
+
 if __name__ == '__main__':
   unittest.main()

test/null/test_uop_graph.py

@@ -423,10 +423,12 @@ class TestUOpGraph(unittest.TestCase):
     d0 = UOp(Ops.PARAM, dtypes.long.ptr(), (), 0)
     ld = d0.index(ridx0.valid(ridx0<50))
     w = (ridx0<50).where(ld, 5)
-    uops = to_uops_list([w])
+    # prevent ridx0 from being shrunk
+    red = UOp(Ops.REDUCE, dtypes.long, (ridx0.cast(dtypes.long), ridx0), Ops.ADD)
+    uops = to_uops_list([w, red])
     for u in uops:
       assert u.op is not Ops.WHERE
-      if u.op is Ops.LOAD: assert u.src[1].arg==5
+      if u.op is Ops.LOAD and u.src[0].src[0].op is Ops.PARAM: assert u.src[1].arg==5
 
   def test_where_on_gated_load_folds_swapped_branches(self):
     ridx0 = UOp.range(100, 0)
@@ -444,10 +446,12 @@ class TestUOpGraph(unittest.TestCase):
     gate_idx = ridx0.valid((ridx0<50))
     ld = d0.index(gate_idx).cast(dtypes.float)
     w = (ridx0<50).where(ld, 5.0)
-    uops = to_uops_list([w])
+    # prevent ridx0 from being shrunk
+    red = UOp(Ops.REDUCE, dtypes.long, (ridx0.cast(dtypes.long), ridx0), Ops.ADD)
+    uops = to_uops_list([w, red])
     for u in uops:
       assert u.op is not Ops.WHERE
-      if u.op is Ops.LOAD: assert u.src[1].arg == 5
+      if u.op is Ops.LOAD and u.src[0].src[0].op is Ops.PARAM: assert u.src[1].arg == 5
 
   def test_where_in_store_becomes_gate(self):
     ridx0 = UOp.range(100, 0)

test/null/test_uop_symbolic.py

@@ -286,6 +286,11 @@ class TestSymbolic(unittest.TestCase):
                               "(((z+(x*-1))+(y*-1))+7)")
     self.helper_test_variable((10+12*Variable("x",0,2)+Variable("y", 0, 4)%3)%13, 8, 12, "(((x*-1)+(y%3))+10)")
 
+  def test_mod_congruence_tied_remainder(self):
+    # when f%c == c/2, both r and r-c have equal abs — try both signs
+    self.helper_test_variable((3+2*Variable("x",0,1)+3*Variable("y",0,1))%4, 0, 3, "((x*-2)+(y*-1)+3)")
+    self.helper_test_variable((3+6*Variable("x",0,1)+7*Variable("y",0,1))%4, 0, 3, "((x*-2)+(y*-1)+3)")
+
   def test_div_congruence(self):
     self.helper_test_variable((3+3*Variable("a",0,3))//4, 0, 3, "a")
     self.helper_test_variable((18+17*Variable("a",0,2)+17)//18, 1, 3, "(a+1)")
@@ -703,6 +708,24 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((31*b+1)//18, 0, 172, "(((b*13)+1)//18+b)")
     self.helper_test_variable((19*b+3)//7, 0, 271, "(((b*5)+3)//7+(b*2))")
 
+  def test_gcd_with_remainder(self):
+    # gcd_with_remainder: factor GCD out of non-constant terms and denominator
+    a = Variable("a", 0, 2)
+    self.helper_test_variable((a*4)//6, 0, 1, "(a*2//3)")
+    self.helper_test_variable((a*4+1)//6, 0, 1, "(a*2//3)")
+    self.helper_test_variable((a*4+2)//6, 0, 1, "((a*2+1)//3)")
+    self.helper_test_variable((a*4+3)//6, 0, 1, "((a*2+1)//3)")
+    self.helper_test_variable((a*4)%6, 0, 4, "(a*2%3*2)")
+    self.helper_test_variable((a*4+1)%6, 1, 5, "(a*2%3*2+1)")
+    self.helper_test_variable((a*4+2)%6, 0, 4, "((a*2+1)%3*2)")
+    self.helper_test_variable((a*4+3)%6, 1, 5, "((a*2+1)%3*2+1)")
+
+  def test_div_by_factor_tie_break(self):
+    a = Variable("a", 0, 1)
+    b = Variable("b", 0, 1)
+    with Context(CORRECT_DIVMOD_FOLDING=1):
+      self.helper_test_variable((a*2+b*3+2)//6, 0, 1, "((a+b+1)//3)")
+
   def test_div_mod_recombine_large_coeff(self):
     # recombine must work even when coeff > divisor: both mod and div reduce the coeff the same way
     b = Variable("b", 0, 100)
@@ -710,6 +733,87 @@ class TestSymbolic(unittest.TestCase):
     a = Variable("a", 0, 10)
     self.helper_test_variable((25*a+3)%10 + ((25*a+3)//10)*10, 3, 253, "((a*25)+3)")
 
+  def test_mod_nest_by_factor(self):
+    # (a*f+b) % (f*k) = (a%k)*f + b when 0<=b<f — mirrors nest_div_by_factor for MOD
+    gidx0 = Variable("gidx0", 0, 15)
+    lidx0 = Variable("lidx0", 0, 3)
+    # f=4, k=2, c=8: (gidx0*4+lidx0)%8 = (gidx0%2)*4 + lidx0
+    self.helper_test_variable((gidx0*4+lidx0)%8, 0, 7, "(lidx0+gidx0%2*4)")
+    # f=2, k=4: (gidx0*2+lidx0)%8 where lidx0 in [0,1]
+    lidx1 = Variable("lidx1", 0, 1)
+    self.helper_test_variable((gidx0*2+lidx1)%8, 0, 7, "(lidx1+gidx0%4*2)")
+    # f=3, k=3: (a*3+b)%9 where b in [0,2]
+    a = Variable("a", 0, 10)
+    b = Variable("b", 0, 2)
+    self.helper_test_variable((a*3+b)%9, 0, 8, "(b+a%3*3)")
+
+  def test_mod_nest_by_factor_with_const(self):
+    # nest_by_factor MOD with non-zero constant offset: (a*f+b+const) % (f*k) = (a%k)*f + b + const when 0<=b+const<f
+    a = Variable("a", 0, 7)
+    b = Variable("b", 0, 1)
+    # f=4, k=2, const=2: (a*4+b+2)%8 = (a%2)*4 + b + 2
+    self.helper_test_variable((a*4+b+2)%8, 2, 7, "(b+a%2*4+2)")
+    # f=6, k=2, const=3: (a*6+b+3)%12 = (a%2)*6 + b + 3
+    b2 = Variable("b", 0, 2)
+    self.helper_test_variable((a*6+b2+3)%12, 3, 11, "(b+a%2*6+3)")
+    # f=3, k=2, const=1: (a*3+b+1)%6 = (a%2)*3 + b + 1
+    self.helper_test_variable((a*3+b+1)%6, 1, 5, "(b+a%2*3+1)")
+
+  def test_div_nest_by_factor_with_const(self):
+    # nest_by_factor IDIV: (160*a + 5*b + 4*c + K) // 60 should pick div=5 (clean) over div=4 (dirty)
+    a = Variable("a", 0, 2)
+    b = Variable("b", 0, 31)
+    c = Variable("c", 0, 1)
+    self.helper_test_variable((160*a + 5*b + 4*c) // 60, 0, 7, "(a*2+(b+a*8)//12)")
+    self.helper_test_variable((160*a + 5*b + 4*c + 1) // 60, 0, 8, "(a*2+(b+c+a*8)//12)")
+    self.helper_test_variable((160*a + 5*b + 4*c + 2) // 60, 0, 8, "(a*2+(b+c+a*8)//12)")
+    self.helper_test_variable((160*a + 5*b + 4*c + 3) // 60, 0, 8, "(a*2+(b+c+a*8)//12)")
+    self.helper_test_variable((160*a + 5*b + 4*c + 59) // 60, 0, 8, "(a*2+(b+c+a*8+11)//12)")
+
+  def test_div_mod_recombine_after_nesting(self):
+    # when nest_div_by_factor simplifies the div, the mod must also nest so recombine can fire
+    gidx0 = Variable("gidx0", 0, 15)
+    lidx0 = Variable("lidx0", 0, 3)
+    x = gidx0*4+lidx0
+    # div nests: x//8 -> gidx0//2, mod nests: x%8 -> (gidx0%2)*4+lidx0, then recombine gives x back
+    self.helper_test_variable((x//8)*8 + x%8, 0, 63, "(lidx0+gidx0*4)")
+    # with a scaling factor: recombine gives x*2
+    self.helper_test_variable((x//8)*16 + (x%8)*2, 0, 126, "(gidx0*8+lidx0*2)")
+    # two variables with different factors
+    a = Variable("a", 0, 7)
+    b = Variable("b", 0, 1)
+    y = a*6+b
+    # div nests: y//12 -> a//2, mod nests: y%12 -> (a%2)*6+b, recombine
+    self.helper_test_variable((y//12)*12 + y%12, 0, 43, "(b+a*6)")
+
+  def test_div_mod_recombine_after_asymmetric_fold(self):
+    a = Variable("a", 0, 7)
+    b = Variable("b", 0, 14)
+    x = a*15+b
+    # TODO: expected "(b+a*15)"
+    self.helper_test_variable((x//10)*10 + x%10, 0, 119, "(a*10+(a+b//5)//2*10+(b+a*5)%10)")
+    self.helper_test_variable((x//10)*2 + (x//5)%2, 0, 23, "(a*3+b//5)")
+
+  def test_div_mod_recombine_in_additive_sum(self):
+    x = Variable("x", 0, 31)
+    y = Variable("y", 0, 5)
+    # recombine should work inside larger additive sums, not just in the two special y+... tree shapes
+    self.helper_test_variable((x//8)*4 + y + (x//2)%4, 0, 20, "(y+x//2)")
+    self.helper_test_variable(y + (x//8)*4 + (x//2)%4, 0, 20, "(y+x//2)")
+
+  def test_div_mod_recompose_low_order_remainder(self):
+    x = Variable("x", 0, 127)
+    self.helper_test_variable((x//2)%4*2 + x%2, 0, 7, "(x%8)")
+
+  def test_reshape_index_roundtrip(self):
+    # simulate reshape index decompose then recompose — the core pattern this enables
+    # (8,8) decomposed for (16,4): combined=r0*8+r1, div and mod by 4
+    r0 = Variable("r0", 0, 7)
+    r1 = Variable("r1", 0, 7)
+    combined = r0*8+r1
+    src_idx = (combined//4)*4 + combined%4
+    self.helper_test_variable(src_idx, 0, 63, "(r1+r0*8)")
+
   def test_gated_load(self):
     idx = Variable("idx", 0, 24)
     self.helper_test_variable(idx//4, 0, 6, "(idx//4)")
@@ -856,6 +960,12 @@ class TestSymbolic(unittest.TestCase):
     self.assertIn((a.cast(dtypes.long)+b.cast(dtypes.long)).render(), "(long)((a+b))")
     self.assertIn((a.cast(dtypes.long)*b.cast(dtypes.long)).render(), "(long)((a*b))")
 
+  def test_nested_mod_negative_range(self):
+    # (x%(k*c))%c = x%c holds for cmod regardless of signs since sign(x%(k*c)) = sign(x)
+    x = Variable("x", 0, 1575)
+    self.helper_test_variable(((x + (-1064)) % 512) % 4, -3, 3, "((x+-1064)%4)")
+    self.helper_test_variable(((x + (-1064)) % 512) % 128, -127, 127, "((x+-1064)%128)")
+
 class TestSymbolicNumeric(unittest.TestCase):
   def helper_test_numeric(self, f):
     MIN, MAX = 0, 10

test/null/test_uop_vmin_vmax.py

@@ -64,8 +64,8 @@ class TestVminVmaxProperties(unittest.TestCase):
 
     # negative mask: x & -1 could be anything since -1 has all bits set
     uop = x & -1
-    self.assertEqual(uop.vmin, dtypes.min(dtypes.int32))
-    self.assertEqual(uop.vmax, dtypes.max(dtypes.int32))
+    self.assertEqual(uop.vmin, dtypes.int32.min)
+    self.assertEqual(uop.vmax, dtypes.int32.max)
 
   def test_vmin_vmax_multiplication_with_variable(self):
     # vmin and vmax for multiplication with a variable
@@ -136,8 +136,8 @@ class TestVminVmaxProperties(unittest.TestCase):
     self.assertEqual(x_bool.vmin, False)
     self.assertEqual(x_bool.vmax, True)
     x_uint = x.cast(dtypes.uint)
-    self.assertEqual(x_uint.vmin, dtypes.min(dtypes.uint))
-    self.assertEqual(x_uint.vmax, dtypes.max(dtypes.uint))
+    self.assertEqual(x_uint.vmin, dtypes.uint.min)
+    self.assertEqual(x_uint.vmax, dtypes.uint.max)
 
   def test_vmin_vmax_invalid(self):
     i = UOp.invalid()

test/null/test_viz.py

@@ -411,7 +411,7 @@ class TestVizProfiler(BaseTestViz):
     j = load_profile(prof)
     event = j['layout']['NV:SDMA:0']['events'][0]
     gbs = sz/(dur*1e-6)*1e-9
-    self.assertTrue(event['fmt'].startswith(f"{gbs:.0f} GB/s"))
+    self.assertEqual(event['fmt'], f"{gbs:.0f} GB/s")
 
   def test_graph(self):
     prof = [ProfileDeviceEvent(device='NV', tdiff=decimal.Decimal(-1000)),
@@ -453,9 +453,7 @@ class TestVizProfiler(BaseTestViz):
     j = load_profile(prof)
     sdma_events = j['layout']['NV:1:SDMA:0']['events']
     gbs = sz/(dur*1e-6)*1e-9
-    timing_txt, trace_txt = sdma_events[0]['fmt'].split("\nTB:")
-    self.assertEqual(timing_txt, f"{gbs:.0f} GB/s")
-    self.assertEqual(json.loads(trace_txt)[0][0], __file__)
+    self.assertEqual(sdma_events[0]['fmt'], f"{gbs:.0f} GB/s")
 
   def test_block_ordering(self):
     prof = [ProfileDeviceEvent(device='NV', tdiff=decimal.Decimal(-1000)),
@@ -511,9 +509,15 @@ class TestVizProfiler(BaseTestViz):
 
   def test_calltrace(self):
     def fxn(): return Tensor.empty(10).mul(2).realize()
-    fxn()
-    trace = get_viz_list()[0]["steps"][0]["trace"]
-    assert any(fxn.__code__.co_filename == f and fxn.__code__.co_firstlineno == l for f,l,*_ in trace), str(trace)
+    with cpu_profile(TracingKey("test_fxn"), "CUSTOM"):
+      fxn()
+    codegen_trace = get_viz_list()[0]["steps"][0]["trace"]
+    assert any(fxn.__code__.co_filename == f and fxn.__code__.co_firstlineno == l for f,l,*_ in codegen_trace), str(codegen_trace)
+    profile_ret = load_profile(cpu_events)
+    e = profile_ret["layout"]["CUSTOM"]["events"][0]
+    self.assertEqual(e["name"], "test_fxn")
+    runtime_trace = json.loads(e["fmt"].replace("TB:", ""))
+    assert any(fxn.__code__.co_filename == f and fxn.__code__.co_firstlineno+1 == l for f,l,*_ in runtime_trace), str(runtime_trace)
 
   # can pack up to 1hr 11 min of trace events
   def test_trace_duration(self):

test/testextra/test_cfg_viz.py

@@ -58,7 +58,8 @@ class TestCfg(unittest.TestCase):
     k.emit(s_endpgm())
     k.emit(s_code_end())
     ei = run_asm("diamond", k)
-    cfg = amdgpu_cfg(ei.prg.p.lib, self.arch)["data"]
+    ret = amdgpu_cfg(ei.prg.p.lib, self.arch)
+    cfg = ret["data"]
     self.assertEqual(len(cfg["blocks"]), 5)
     edge_count = sum(len(v) for v in cfg["paths"].values())
     self.assertEqual(edge_count, 5)
@@ -69,8 +70,11 @@ class TestCfg(unittest.TestCase):
     self.assertEqual(len(references["r0"]), 2)
     insts = [cfg["pc_tokens"][pc][0]["st"] for pc in references["r0"]]
     self.assertEqual(insts, ['s_mov_b32', 's_cmp_eq_u64'])
-    end_block_content = "\n".join(" ".join(t["st"] for t in cfg["pc_tokens"][pc]) for pc in list(cfg["blocks"].values())[-1])
-    self.assertEqual(end_block_content, "s_endpgm\ns_code_end (217x)")
+    end_block = [" ".join(t["st"] for t in cfg["pc_tokens"][pc]) for pc in list(cfg["blocks"].values())[-1]]
+    code_line = ret["src"].splitlines()[-1]
+    self.assertEqual(len(end_block), 2)
+    for st in [end_block[-1], code_line]:
+      assert st.startswith("s_code_end") and st.endswith("x)"), st
 
   def test_loop(self):
     k = Kernel(arch=Device["AMD"].arch)

test/unit/test_assign.py

@@ -272,14 +272,19 @@ class TestAssign(unittest.TestCase):
 
   def test_assign_after(self):
     t = Tensor.zeros(10).contiguous().realize()
-    t.uop = t.uop.after(t.uop.assign((t+1).uop))
+    t.uop = t.uop.after(t.uop.store((t+1).uop))
     np.testing.assert_allclose(t.numpy(), [1.,1.,1.,1.,1.,1.,1.,1.,1.,1.])
 
   def test_assign_after_partial(self):
     t = Tensor.zeros(10).contiguous().realize()
-    t.uop = t.uop.after(t[:5].uop.assign(Tensor.ones(5).uop))
+    t.uop = t.uop.after(t[:5].uop.after(t[:5].uop.store(Tensor.ones(5).uop)))
     np.testing.assert_allclose(t.numpy(), [1.,1.,1.,1.,1.,0.,0.,0.,0.,0.])
 
+  def test_assign_after_target_chain(self):
+    t = Tensor.arange(16).reshape(4, 4).permute(1, 0).contiguous()
+    t.assign(t + 100)
+    np.testing.assert_equal(t.numpy(), [[100, 104, 108, 112], [101, 105, 109, 113], [102, 106, 110, 114], [103, 107, 111, 115]])
+
   def test_assign_contiguous(self):
     b = Tensor.arange(16).reshape(4,4).contiguous().realize()
     a = (Tensor.arange(16).reshape(4,4).contiguous().realize() + 1)
@@ -911,5 +916,36 @@ class TestAssignToUnrealizedView(unittest.TestCase):
       # TODO: broken now, silently dropped
       self.assertEqual(c.tolist(), [[5,5],[5,5]])
 
+class TestPartialAssignToSharedBuffer(unittest.TestCase):
+  def test_five_slices(self):
+    big = Tensor.zeros(50).contiguous().realize()
+    views = [big[i*10:(i+1)*10].reshape(2, 5) for i in range(5)]
+    for v in views: v.assign(v + 1)
+    Tensor.realize(*views)
+    for v in views:
+      np.testing.assert_allclose(v.numpy(), np.ones((2, 5)))
+
+  def test_many_slices(self):
+    n_params = 10
+    big = Tensor.zeros(n_params * 12).contiguous().realize()
+    grads = [big[i*12:(i+1)*12].reshape(3, 4) for i in range(n_params)]
+    for g in grads: g.assign(g + 1)
+    Tensor.realize(*grads)
+    for g in grads:
+      np.testing.assert_allclose(g.numpy(), np.ones((3, 4)))
+
+  def test_mixed_shapes(self):
+    big = Tensor.zeros(100).contiguous().realize()
+    shapes = [(3, 4), (4, 6), (6, 4), (2, 5), (4, 3)]
+    pos, views = 0, []
+    for s in shapes:
+      n = s[0] * s[1]
+      views.append(big[pos:pos+n].reshape(*s))
+      pos += n
+    for v in views: v.assign(v + 1)
+    Tensor.realize(*views)
+    for v, s in zip(views, shapes):
+      np.testing.assert_allclose(v.numpy(), np.ones(s))
+
 if __name__ == "__main__":
   unittest.main()

test/unit/test_call.py

@@ -237,5 +237,116 @@ class TestCallSchedule(unittest.TestCase):
     self.assertIsInstance(out.shape[0], UOp)
     np.testing.assert_allclose(out[:5].numpy(), (np.arange(16*4).reshape(16, 4)[:5] * 2 + 1).astype(np.float32))
 
+  def test_precompile_multi_sharded(self):
+    @function(precompile=True)
+    def f(x:Tensor) -> Tensor: return x + 1
+    devs = ("CPU:0", "CPU:1")
+    a = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=0)
+    out = f(a) + 2
+    np.testing.assert_allclose(out.numpy(), np.arange(8, dtype=np.float32).reshape(4, 2) + 3)
+
+class TestCallMultiSharded(unittest.TestCase):
+  # TODO: multi-output + sharded needs per-device CALL execution, which requires reworking how MULTI propagates through TUPLE bodies
+  def test_tuple_sharded(self):
+    """multi-output function with sharded input"""
+    devs = ("CPU:0", "CPU:1")
+    @function
+    def f(x:Tensor): return (x + 1, x * 2)
+    a = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=0)
+    t1, t2 = f(a)
+    ref = np.arange(8, dtype=np.float32).reshape(4, 2)
+    np.testing.assert_allclose(t1.numpy(), ref + 1)
+    np.testing.assert_allclose(t2.numpy(), ref * 2)
+
+  def test_tuple_sharded_precompile(self):
+    """multi-output precompiled function with sharded input"""
+    devs = ("CPU:0", "CPU:1")
+    @function(precompile=True)
+    def f(x:Tensor): return (x + 1, x * 2)
+    a = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=0)
+    t1, t2 = f(a)
+    ref = np.arange(8, dtype=np.float32).reshape(4, 2)
+    np.testing.assert_allclose(t1.numpy(), ref + 1)
+    np.testing.assert_allclose(t2.numpy(), ref * 2)
+
+  def test_tuple_sharded_different_axis(self):
+    """multi-output function where outputs have different sharding: one reduces on sharded axis, one doesn't"""
+    devs = ("CPU:0", "CPU:1")
+    @function
+    def f(x:Tensor): return (x.sum(axis=0), x.sum(axis=1))
+    a = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=0)
+    t1, t2 = f(a)
+    ref = np.arange(8, dtype=np.float32).reshape(4, 2)
+    np.testing.assert_allclose(t1.numpy(), ref.sum(axis=0))
+    np.testing.assert_allclose(t2.numpy(), ref.sum(axis=1))
+
+  def test_tuple_sharded_different_ops(self):
+    """multi-output function with different operations per output"""
+    devs = ("CPU:0", "CPU:1")
+    @function
+    def f(x:Tensor, y:Tensor): return (x + y, x * y)
+    a = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=0)
+    b = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=0) + 1
+    t1, t2 = f(a, b)
+    ref_a = np.arange(8, dtype=np.float32).reshape(4, 2)
+    ref_b = ref_a + 1
+    np.testing.assert_allclose(t1.numpy(), ref_a + ref_b)
+    np.testing.assert_allclose(t2.numpy(), ref_a * ref_b)
+
+  def test_tuple_sharded_mixed_use(self):
+    """multi-output sharded results used in further computation"""
+    devs = ("CPU:0", "CPU:1")
+    @function
+    def f(x:Tensor): return (x + 1, x * 2)
+    a = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=0)
+    t1, t2 = f(a)
+    out = (t1 + t2).sum()
+    ref = np.arange(8, dtype=np.float32).reshape(4, 2)
+    np.testing.assert_allclose(out.numpy(), ((ref + 1) + (ref * 2)).sum())
+
+  def test_tuple_sharded_outputs_different_axis(self):
+    """multi-output function where the two outputs are sharded on different axes"""
+    devs = ("CPU:0", "CPU:1")
+    @function
+    def f(x:Tensor, y:Tensor): return (x + 1, y + 2)
+    a = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=0)
+    b = Tensor.arange(8).reshape(4, 2).float().shard(devs, axis=1)
+    t1, t2 = f(a, b)
+    ref_a = np.arange(8, dtype=np.float32).reshape(4, 2)
+    ref_b = np.arange(8, dtype=np.float32).reshape(4, 2)
+    np.testing.assert_allclose(t1.numpy(), ref_a + 1)
+    np.testing.assert_allclose(t2.numpy(), ref_b + 2)
+
+  def test_call_reduce_sharded(self):
+    devs = ("CPU:0", "CPU:1")
+    a = Tensor.ones(10, 10).shard(devs, axis=0)
+    Tensor.realize(a)
+    c = Tensor.call(a, fxn=a.as_param(0).sum(axis=0))
+    np.testing.assert_equal(c.numpy(), 10 * np.ones(10))
+
+  def test_call_reduce_sharded_mixed_args(self):
+    devs = ("CPU:0", "CPU:1")
+    a = Tensor.ones(10, 10).shard(devs, axis=0)
+    b = Tensor.ones(10).shard(devs, axis=None)
+    Tensor.realize(a, b)
+    c = Tensor.call(a, b, fxn=a.as_param(0).sum(axis=0) + b.as_param(1))
+    np.testing.assert_equal(c.numpy(), 11 * np.ones(10))
+
+  def test_call_reduce_sharded_backward(self):
+    devs = ("CPU:0", "CPU:1")
+    a = Tensor.randn(10, 10, requires_grad=True).shard(devs, axis=0)
+    b = Tensor.randn(10, 10, requires_grad=True).shard(devs, axis=0)
+    Tensor.realize(a, b)
+
+    def grad_fxn(grad, call):
+      a_arg, b_arg = call.src[1], call.src[2]
+      return (grad.expand(a_arg.shape) * b_arg, grad.expand(b_arg.shape) * a_arg)
+
+    body = (a.as_param(0) * b.as_param(1)).sum(axis=0)
+    c = Tensor.call(a, b, fxn=body, grad_fxn=grad_fxn)
+    c.sum().backward()
+    np.testing.assert_allclose(a.grad.numpy(), b.numpy(), rtol=1e-5)
+    np.testing.assert_allclose(b.grad.numpy(), a.numpy(), rtol=1e-5)
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_dtype_spec.py

@@ -17,13 +17,15 @@ dtype_floats = [dt for dt in core_dtypes if dtypes.is_float(dt) and is_dtype_sup
 
 FP8E4M3_MAX = 448.0
 FP8E5M2_MAX = 57344.0
+FP8E4M3FNUZ_MAX = 240.0
+FP8E5M2FNUZ_MAX = 57344.0
 
 def _assert_eq(tensor:Tensor, target_dtype:DType, target, tol_target_dtype:float=1e-7):
   if DEBUG >= 2: print(tensor.numpy())
   try:
     assert tensor.dtype == target_dtype
     np.testing.assert_allclose(tensor.numpy(), target, rtol={dtypes.float16:1e-3, dtypes.bfloat16:1e-2,
-      dtypes.fp8e4m3:1e-1, dtypes.fp8e5m2:5e-1}.get(target_dtype, tol_target_dtype))
+      dtypes.fp8e4m3:1e-1, dtypes.fp8e5m2:5e-1, dtypes.fp8e4m3fnuz:1e-1, dtypes.fp8e5m2fnuz:5e-1}.get(target_dtype, tol_target_dtype))
 
   except AssertionError as e:
     raise AssertionError(f"\ntensor {tensor.numpy()} dtype {tensor.dtype} does not match target {target} with dtype {target_dtype}") from e

test/unit/test_function.py

@@ -1,7 +1,7 @@
 import numpy as np
 import unittest
 from tinygrad.function import function
-from tinygrad import Tensor
+from tinygrad import Tensor, GlobalCounters
 from tinygrad.uop.ops import UOp
 
 class TestFunction(unittest.TestCase):
@@ -165,13 +165,13 @@ class TestFunction(unittest.TestCase):
   def test_name(self):
     @function
     def f(a:Tensor) -> Tensor: return a + 1
-    assert f(Tensor([1])).uop.arg.name.endswith("f")
+    assert f(Tensor([1])).uop.src[0].arg.name.endswith("f")
 
   def test_method_name(self):
     class Foo:
       @function
       def __call__(self, x:Tensor) -> Tensor: return x + 1
-    assert Foo()(Tensor([1])).uop.arg.name.endswith("Foo.__call__")
+    assert Foo()(Tensor([1])).uop.src[0].arg.name.endswith("Foo.__call__")
 
   def test_callable_instance(self):
     class Foo:
@@ -180,7 +180,7 @@ class TestFunction(unittest.TestCase):
     foo = Foo()
     f = function(foo)
     np.testing.assert_equal(f(Tensor([1,2,3])).numpy(), [11,22,33])
-    assert f(Tensor([1,2,3])).uop.arg.name.endswith("Foo")
+    assert f(Tensor([1,2,3])).uop.src[0].arg.name.endswith("Foo")
 
   def test_iadd(self):
     @function
@@ -213,6 +213,17 @@ class TestFunction(unittest.TestCase):
     np.testing.assert_equal(a.numpy(), [11,21,31])  # TODO: should be [1,2,3]
     np.testing.assert_equal(b.numpy(), [10,20,30])
 
+  def test_view_assign_explicit_buffer(self):
+    """view assign on an explicit param's buffer should not create implicit inputs."""
+    class State:
+      def __init__(self): self.buf = Tensor.zeros(2, 4).contiguous().realize()
+      @function(allow_implicit=False)
+      def __call__(self, x:Tensor) -> Tensor:
+        self.buf[:, 0:2].assign(x)
+        return self.buf[:, 0:2]
+    s = State()
+    np.testing.assert_equal(s(Tensor([[5., 6.], [7., 8.]])).numpy(), [[5., 6.], [7., 8.]])
+
   @unittest.expectedFailure
   def test_assign_slice(self):
     @function
@@ -335,5 +346,69 @@ class TestFunctionMulti(unittest.TestCase):
       f(x).sum().backward()
       np.testing.assert_allclose(x.grad.numpy(), expected)
 
+class TestFunctionTuple(unittest.TestCase):
+  def test_tuple(self, precompile=False):
+    x = Tensor.ones(3).contiguous()
+    @function(precompile=precompile)
+    def f(t:Tensor): return (t+1, t+2)
+    t1, t2 = f(x)
+    t1.realize(t2)
+    print(t1.tolist(), t2.tolist())
+    assert t1.tolist() == [2,2,2]
+    assert t2.tolist() == [3,3,3]
+  def test_tuple_precompile(self): self.test_tuple(True)
+
+  def test_grad_tuple(self, precompile=False):
+    x = Tensor.ones(3, requires_grad=True).contiguous()
+    y = Tensor.ones(3, requires_grad=True).contiguous()
+    @function(precompile=precompile)
+    def f(u1:Tensor, u2:Tensor): return (u1+1, u2+2)
+    t1, t2 = f(x,y)
+    (t1+t2).sum().backward()
+    x.grad.realize(y.grad)
+  def test_grad_tuple_precompile(self): self.test_grad_tuple(True)
+
+  def test_grad_fxn_tuple(self):
+    # grad_fxn for tuple: receives one gradient per output as positional args
+    def grad_fxn(d_out0:UOp, d_out1:UOp, call:UOp):
+      # f(u1, u2) = (u1+1, u2+2)
+      # df/du1 = d_out0, df/du2 = d_out1
+      return (d_out0, d_out1)
+
+    x = Tensor.ones(3, requires_grad=True).contiguous()
+    y = Tensor.ones(3, requires_grad=True).contiguous()
+    @function(grad_fxn=grad_fxn)
+    def f(u1:Tensor, u2:Tensor): return (u1+1, u2+2)
+    t1, t2 = f(x, y)
+    (t1+t2).sum().backward()
+    np.testing.assert_allclose(x.grad.numpy(), [1., 1., 1.])
+    np.testing.assert_allclose(y.grad.numpy(), [1., 1., 1.])
+
+class TestFunctionGrad(unittest.TestCase):
+  def test_function_grad_ops(self, precompile=False, precompile_backward=False):
+    N = 64
+    x = Tensor.ones(N,N).contiguous()
+    w1 = Tensor.ones(N,N, requires_grad=True).contiguous()
+    w2 = Tensor.ones(N,N, requires_grad=True).contiguous()
+    w3 = Tensor.ones(N,N, requires_grad=True).contiguous()
+    ref = Tensor.ones(N,N).contiguous()
+    Tensor.realize(x, w1, w2, w3, ref)
+    @function(precompile=precompile, precompile_backward=precompile_backward)
+    def f(x, w1, w2, w3) -> tuple[Tensor, ...]:
+      p1 = x@w1
+      p2 = p1@w2
+      p3 = p2@w3
+      return p1, p2, p3, p3.contiguous()
+    ret = f(x, w1, w2, w3)[-1]
+    loss = (ret-ref).square().mean().backward()
+    print("RESET")
+    GlobalCounters.reset()
+    loss.realize(w1.grad, w2.grad, w3.grad)
+    print(GlobalCounters.global_ops, GlobalCounters.global_mem)
+    self.assertLessEqual(GlobalCounters.global_ops, 4739073)
+  def test_function_grad_ops_precompile(self): self.test_function_grad_ops(precompile=True)
+  def test_function_grad_ops_precompile_backward(self):
+    self.test_function_grad_ops(precompile=True, precompile_backward=True)
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_gguf.py

@@ -26,7 +26,7 @@ class TestGGUF(unittest.TestCase):
     # codes 0-7 = [0, 1, 2, 3, 4, 6, 8, 12], codes 8-15 are their negatives
     block = np.array([0x80] + list(range(16)), dtype=np.uint8)  # E=128, nibbles 0-15 in low, zeros in high
     expected = np.array([0., 1., 2., 3., 4., 6., 8., 12., -0., -1., -2., -3., -4., -6., -8., -12.] + [0.]*16, dtype=np.float32)
-    np.testing.assert_equal(ggml_data_to_tensor(Tensor(block), 32, 39).numpy().flatten(), expected)
+    np.testing.assert_equal(ggml_data_to_tensor(Tensor(block), 32, GGMLQuantizationType.MXFP4.value).numpy().flatten(), expected)
 
   def test_dequantization_q4_0(self): self._test_dequantization(GGMLQuantizationType.Q4_0)
   def test_dequantization_q4_1(self): self._test_dequantization(GGMLQuantizationType.Q4_1)
@@ -34,9 +34,8 @@ class TestGGUF(unittest.TestCase):
   def test_dequantization_q4_k(self): self._test_dequantization(GGMLQuantizationType.Q4_K)
   def test_dequantization_q5_k(self): self._test_dequantization(GGMLQuantizationType.Q5_K)
   def test_dequantization_q6_k(self): self._test_dequantization(GGMLQuantizationType.Q6_K)
-  def test_dequantization_mxfp4(self):
-    MXFP4 = 39
-
+  def test_dequantization_mxfp4(self): self._test_dequantization(GGMLQuantizationType.MXFP4)
+  def test_dequantization_mxfp4_old(self):
     def encode(nibbles, E):
       packed = [(low & 0xF) | ((high & 0xF) << 4) for low, high in zip(nibbles[:16], nibbles[16:])]
       return np.array([E] + packed, dtype=np.uint8)
@@ -57,12 +56,10 @@ class TestGGUF(unittest.TestCase):
       blocks.append(encode(codes, E))
       expected.extend(decode(c, E) for c in codes)
     tensor = Tensor(np.concatenate(blocks))
-    out = ggml_data_to_tensor(tensor, len(expected), MXFP4)
-    # TODO: should this be exact equal? somehow failed on CI
-    np.testing.assert_allclose(out.numpy(), expected, atol=0.0, rtol=1e-6)
+    out = ggml_data_to_tensor(tensor, len(expected), GGMLQuantizationType.MXFP4.value)
+    np.testing.assert_equal(out.numpy(), expected)
 
   def test_dequantization_mxfp4_block(self):
-    MXFP4 = 39
     # https://gist.github.com/Ananta-Ranganathan/3317b6ed51a3b033e9c2564fafb4e043
     # used the above script to download the first block of blk.0.attn_k_b.weight from
     # https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF/blob/main/GLM-4.7-Flash-MXFP4_MOE.gguf
@@ -77,9 +74,8 @@ class TestGGUF(unittest.TestCase):
                         0.03125000,  0.00000000, 0.06250000,  0.01562500,
                         -0.06250000,  0.00000000, 0.00000000, -0.01562500,
                         0.04687500,  0.00000000, 0.00000000,  0.01562500], dtype=np.float32)
-    out = ggml_data_to_tensor(Tensor(block), 32, MXFP4)
-    # TODO: similar to previous test fails on Mac CI with assert_equal for unclear reason
-    np.testing.assert_allclose(out.numpy(), expected, atol=0.0, rtol=1e-6)
+    out = ggml_data_to_tensor(Tensor(block), 32, GGMLQuantizationType.MXFP4.value)
+    np.testing.assert_equal(out.numpy(), expected)
 
   def test_expected_failure_unknown_type(self):
     with self.assertRaises(ValueError):
@@ -133,6 +129,7 @@ class TestGGUFGEMV(unittest.TestCase):
     if qtype == GGMLQuantizationType.Q8_0: q_data[:, :2] = scales[:, :2]                    # d at offset 0
     elif qtype in (GGMLQuantizationType.Q4_K, GGMLQuantizationType.Q5_K): q_data[:, :4] = scales[:, :4] # d, dmin at offset 0
     elif qtype == GGMLQuantizationType.Q6_K: q_data[:, -2:] = scales[:, :2]                 # d at end
+    elif qtype == GGMLQuantizationType.MXFP4: q_data[:, 0] = rng.integers(120, 136, size=n_blocks, dtype=np.uint8) # constrain byte0
     q_data = q_data.flatten()
     ref = dequantize(q_data, qtype).reshape(rows, cols)
 
@@ -158,6 +155,7 @@ class TestGGUFGEMV(unittest.TestCase):
   def test_gguf_gemv_q4_k(self): self._test_gguf_gemv(GGMLQuantizationType.Q4_K)
   def test_gguf_gemv_q5_k(self): self._test_gguf_gemv(GGMLQuantizationType.Q5_K)
   def test_gguf_gemv_q6_k(self): self._test_gguf_gemv(GGMLQuantizationType.Q6_K)
+  def test_gguf_gemv_mxfp4(self): self._test_gguf_gemv(GGMLQuantizationType.MXFP4)
 
 if __name__ == '__main__':
   unittest.main()

test/unit/test_invalid_tensor.py

@@ -0,0 +1,126 @@
+import unittest
+from tinygrad import Tensor
+from tinygrad.dtype import Invalid, dtypes
+from tinygrad.engine.realize import run_schedule
+
+class TestInvalidTensor(unittest.TestCase):
+  def _invalid_test_helper(self, out, expected):
+    sched = out.schedule()
+    buf = out.uop.buffer
+    buf.allocate()
+    sentinel = memoryview(bytearray(b'\x42' * buf.nbytes))
+    buf.copyin(sentinel)
+    before = buf.as_memoryview().cast(out.dtype.fmt).tolist()
+    run_schedule(sched)
+    ret = buf.as_memoryview().cast(out.dtype.fmt).tolist()
+
+    for i,v in enumerate(expected): self.assertEqual(ret[i], before[i] if v is None else v)
+
+  def test_where_x_invalid(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Invalid)
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+  def test_where_invalid_x(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Invalid, Tensor([1.0, 2.0, 3.0, 4.0]))
+    self._invalid_test_helper(out, [None, None, 3.0, 4.0])
+
+  def test_where_invalid_2d(self):
+    mask = Tensor.arange(6).reshape(2, 3) < 3
+    vals = Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
+    out = mask.where(vals, Invalid)
+    self._invalid_test_helper(out, [1.0, 2.0, 3.0, None, None, None])
+
+  def test_where_invalid_int(self):
+    mask = Tensor.arange(3) < 2
+    out = mask.where(Tensor([10, 20, 30]), Invalid)
+    self._invalid_test_helper(out, [10, 20, None])
+
+  def test_where_invalid_add(self):
+    mask = Tensor.arange(3) < 2
+    mixed = mask.where(Tensor([10.0, 20.0, 30.0]), Invalid)
+    out = mixed + Tensor([1.0, 2.0, 3.0])
+    self._invalid_test_helper(out, [11.0, 22.0, None])
+
+  def test_where_invalid_add_left(self):
+    mask = Tensor.arange(3) < 2
+    mixed = mask.where(Tensor([10.0, 20.0, 30.0]), Invalid)
+    out = Tensor([1.0, 2.0, 3.0]) + mixed
+    self._invalid_test_helper(out, [11.0, 22.0, None])
+
+  def test_where_always_true(self):
+    mask = Tensor.arange(3) < 10
+    out = mask.where(Tensor([10.0, 20.0, 30.0]), Invalid)
+    self._invalid_test_helper(out, [10.0, 20.0, 30.0])
+
+  def test_where_cast(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Invalid).cast(dtypes.int)
+    self._invalid_test_helper(out, [1, 2, None, None])
+
+  def test_where_compare(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Invalid) > 1
+    self._invalid_test_helper(out, [False, True, None, None])
+
+  def test_where_unary(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 4.0, 9.0, 16.0]), Invalid).sqrt()
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+  def test_where_where(self):
+    mask1 = Tensor.arange(4) < 2
+    mask2 = Tensor.arange(4) > 0
+    out = mask2.where(mask1.where(Tensor([1.0, 2.0, 3.0, 4.0]), Invalid), Invalid)
+    self._invalid_test_helper(out, [None, 2.0, None, None])
+
+  def test_where_reduce_always_true(self):
+    mask = Tensor.arange(4) < 9
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Invalid).sum()
+    self._invalid_test_helper(out, [10.0])
+
+  def test_invalid_unary(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Tensor.full((4,), Invalid, dtype=dtypes.float).sqrt())
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+  def test_invalid_binary(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Tensor.full((4,), Invalid, dtype=dtypes.float) + 2)
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+  def test_invalid_binary_left(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), 2 + Tensor.full((4,), Invalid, dtype=dtypes.float))
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+  def test_invalid_reshape(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Invalid).reshape(2,2)
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+  def test_invalid_cast(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Tensor.full((4,), Invalid, dtype=dtypes.int).cast(dtypes.float))
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+  def test_invalid_bitcast(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Tensor.full((4,), Invalid, dtype=dtypes.int).bitcast(dtypes.float))
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+  def test_where_bitcast(self):
+    mask = Tensor.arange(4) < 2
+    out = mask.where(Tensor([1.0, 2.0, 3.0, 4.0]), Tensor.full((4,), Invalid, dtype=dtypes.int)).bitcast(dtypes.int)
+    self._invalid_test_helper(out, [0x3f800000, 0x40000000, None, None])
+
+  # tensor indexing uses reduce, so the entire result becomes invalid
+  @unittest.expectedFailure
+  def test_tensor_index(self):
+    idx = (Tensor.arange(4) < 2).where(Tensor([0, 1, 2, 3]), Invalid)
+    out = Tensor([1.0, 2.0, 3.0, 4.0])[idx]
+    self._invalid_test_helper(out, [1.0, 2.0, None, None])
+
+if __name__ == '__main__':
+  unittest.main()

test/unit/test_llm_moe.py

@@ -49,5 +49,26 @@ class TestMoEFeedForward(unittest.TestCase):
     expected = 1 + (Tensor([1.0]).silu().item() + Tensor([3.0]).silu().item()) / 2
     np.testing.assert_allclose(out.numpy(), expected, rtol=1e-2)
 
+  def test_moe_feed_forward_norm_topk_prob(self):
+    from tinygrad.apps.llm import TransformerBlock
+    dim, hidden, n_heads = 8, 16, 2
+    num_experts, k = 4, 2
+
+    block = TransformerBlock(dim, hidden, n_heads, n_heads, norm_eps=1e-5, head_dim=dim//n_heads,
+                             rope_theta=10000, max_context=16, num_experts=num_experts, num_experts_per_tok=k)
+    block.norm_topk_prob = True
+
+    block.ffn_gate_exps.weight = Tensor.stack(*[Tensor.eye(hidden, dim) * (i + 1) for i in range(num_experts)])
+    block.ffn_up_exps.weight = Tensor.stack(*[Tensor.eye(hidden, dim) for _ in range(num_experts)])
+    block.ffn_down_exps.weight = Tensor.stack(*[Tensor.eye(dim, hidden) for _ in range(num_experts)])
+    block.ffn_gate_inp.weight = Tensor([[0.1, 0, 0.1, 0]] * dim).T  # equal top-2 experts, but only ~69% mass before renorm
+    block.ffn_norm.weight = Tensor.ones(dim)
+
+    h = Tensor.ones(1, 1, dim)
+    out = block._feed_forward(h)
+
+    expected = 1 + (Tensor([1.0]).silu().item() + Tensor([3.0]).silu().item()) / 2
+    np.testing.assert_allclose(out.numpy()[0, 0, 0], expected, rtol=1e-2)
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_llm_server.py

@@ -83,5 +83,30 @@ class TestTransformerGenerate(unittest.TestCase):
     self.assertEqual(cache_size_after_warmup, len(schedule_cache),
       f"third prompt added {len(schedule_cache) - cache_size_after_warmup} new schedule cache entries (expected 0)")
 
+  def test_chunked_prefill(self):
+    """When prompt > chunk_size, all chunks should be prefill"""
+    from tinygrad.apps.llm import Transformer
+    from tinygrad.uop.ops import resolve
+    model = Transformer(num_blocks=1, dim=64, hidden_dim=128, n_heads=2, n_kv_heads=2,
+                        norm_eps=1e-5, vocab_size=100, head_dim=32, rope_theta=10000.0, max_context=64)
+
+    def get_prefill_flags(tokens, chunk_size):
+      is_prefill = []
+      def mock_call(self, tokens, start_pos):
+        is_prefill.append(resolve(tokens.shape[1] != 1))
+        return Tensor([[42]])
+      with patch.object(Transformer, '__call__', mock_call):
+        gen = model.generate(tokens, chunk_size=chunk_size)
+        for _ in range(3): next(gen)
+      model._cached_tokens = []
+      return is_prefill
+
+    # 8 tokens, chunk_size=4 -> 2 prefill chunks
+    self.assertEqual(get_prefill_flags(list(range(8)), 4), [True, True, False, False])
+    # 9 tokens, chunk_size=4 -> 3 prefill chunks (4+4+1)
+    self.assertEqual(get_prefill_flags(list(range(9)), 4), [True, True, True, False, False])
+    # 4 tokens, chunk_size=4 -> 1 prefill chunk
+    self.assertEqual(get_prefill_flags(list(range(4)), 4), [True, False, False])
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_setitem_schedule.py

@@ -36,8 +36,7 @@ class TestSetitemInto(unittest.TestCase):
     self.assertEqual(GlobalCounters.kernel_count, 0)
     t.realize()
     self.assertEqual(GlobalCounters.kernel_count, 1)
-    # TODO: this can be just 4 if empty goes through is_realized setitem path
-    self.assertEqual(GlobalCounters.global_mem, 4*(3*2+1)) # 3 elements had +1, 1 is assigned directly
+    self.assertEqual(GlobalCounters.global_mem, 4)
     t[1].realize()
     t.realize()
     self.assertEqual(GlobalCounters.kernel_count, 1)

tinygrad/apps/llm.py

@@ -84,7 +84,7 @@ def apply_rope(x:Tensor, freqs_cis:Tensor) -> Tensor:
 
 class TransformerBlock:
   def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int, norm_eps:float, head_dim:int, rope_theta:float,
-               max_context:int=0, qk_norm:int=0, num_experts:int=0, num_experts_per_tok:int=0):
+               max_context:int=0, qk_norm:int=0, num_experts:int=0, num_experts_per_tok:int=0, norm_topk_prob:bool=False):
     self.n_heads      = n_heads
     self.n_kv_heads   = n_kv_heads
     self.head_dim     = head_dim
@@ -107,6 +107,7 @@ class TransformerBlock:
 
     # --- feed-forward (MoE or dense) -------------------------------------
     if num_experts > 0:
+      self.norm_topk_prob = norm_topk_prob
       self.num_experts_per_tok = num_experts_per_tok
       self.ffn_gate_inp = nn.Linear(dim, num_experts, bias=False)  # router
       self.ffn_gate_exps = ExpertWeights(num_experts, dim, hidden_dim)
@@ -117,7 +118,7 @@ class TransformerBlock:
       self.ffn_up      = nn.Linear(dim, hidden_dim, bias=False)
       self.ffn_down    = nn.Linear(hidden_dim, dim, bias=False)
 
-  @function(precompile=bool(getenv("PRECOMPILE", 0)))
+  @function(precompile=bool(getenv("PRECOMPILE", 0)), allow_implicit=False)
   def _attention(self, x:Tensor, start_pos:int|UOp) -> Tensor:
     x_norm = self.attn_norm(x)                       # (B,T,D)
     q, k, v = self.attn_q(x_norm), self.attn_k(x_norm), self.attn_v(x_norm)
@@ -129,19 +130,12 @@ class TransformerBlock:
     v = v.reshape(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2)  # (B,KvH,T,Hd)
     if self.qk_norm == self.head_dim: q, k = self.attn_q_norm(q), self.attn_k_norm(k)
 
-    freqs_cis = precompute_freqs_cis(self.head_dim, self.max_context, self.rope_theta)[start_pos:start_pos+T]
-    q = apply_rope(q, freqs_cis)
-    k = apply_rope(k, freqs_cis)
+    q = apply_rope(q, self.freqs_cis[start_pos:start_pos+T])
+    k = apply_rope(k, self.freqs_cis[start_pos:start_pos+T])
 
-    # TODO: fix assign to behave like this
-    assigned_kv = self.cache_kv.uop.after(self.cache_kv[:, :, :, start_pos:start_pos+T, :].uop.assign(Tensor.stack(k, v).contiguous().uop))
-    tensor_assigned_kv = Tensor(assigned_kv, device=assigned_kv.device)
-    k = tensor_assigned_kv[0, :, :, 0:start_pos+T, :]
-    v = tensor_assigned_kv[1, :, :, 0:start_pos+T, :]
-
-    #self.cache_kv[:, :, :, start_pos:start_pos+T, :].assign(Tensor.stack(k, v))
-    #k = self.cache_kv[0, :, :, 0:start_pos+T, :]
-    #v = self.cache_kv[1, :, :, 0:start_pos+T, :]
+    self.cache_kv[:, :, :, start_pos:start_pos+T, :].assign(Tensor.stack(k, v))
+    k = self.cache_kv[0, :, :, 0:start_pos+T, :]
+    v = self.cache_kv[1, :, :, 0:start_pos+T, :]
 
     # NOTE: this mask is causal_lower_right, not the causal_upper_left generated by is_casual = True
     # TODO: this if statement should be removed and it shouldn't generate extra kernels
@@ -151,12 +145,13 @@ class TransformerBlock:
     attn = self.attn_output(attn)
     return x + attn
 
-  @function(precompile=bool(getenv("PRECOMPILE", 0)))
+  @function(precompile=bool(getenv("PRECOMPILE", 0)), allow_implicit=False)
   def _feed_forward(self, h: Tensor) -> Tensor:
     h_norm = self.ffn_norm(h)
     if hasattr(self, 'ffn_gate_exps'):
       x = h_norm.unsqueeze(2)  # (B, T, 1, D) - add expert dim for broadcasting
       probs, sel = self.ffn_gate_inp(h_norm).softmax(-1).topk(self.num_experts_per_tok)  # (B, T, k) each
+      if self.norm_topk_prob: probs = probs / probs.sum(axis=-1, keepdim=True)
       x_down = self.ffn_down_exps(sel, self.ffn_gate_exps(sel, x).silu() * self.ffn_up_exps(sel, x))  # (B, T, k, D)
       return h + (x_down * probs.unsqueeze(-1)).sum(axis=2)  # (B, T, D)
     # TODO: remove the need for this contiguous
@@ -168,13 +163,14 @@ class TransformerBlock:
       # TODO: how is the dtype of this determined?
       # NOTE: clone is used to promise the creation of a specific buffer
       self.cache_kv = Tensor.zeros(2, x.shape[0], self.n_kv_heads, self.max_context, self.head_dim, device=x.device).clone()
+      self.freqs_cis = precompute_freqs_cis(self.head_dim, self.max_context, self.rope_theta)
     return self._feed_forward(self._attention(x, start_pos)).contiguous()
 
 class Transformer:
   def __init__(self, *, num_blocks, dim, hidden_dim, n_heads, n_kv_heads, norm_eps, vocab_size, head_dim:int, rope_theta:float,
-               max_context:int=0, qk_norm:int=0, num_experts:int=0, num_experts_per_tok:int=0):
+               max_context:int=0, qk_norm:int=0, num_experts:int=0, num_experts_per_tok:int=0, norm_topk_prob:bool=False):
     self.blk = [TransformerBlock(dim, hidden_dim, n_heads, n_kv_heads, norm_eps, head_dim, rope_theta, max_context, qk_norm,
-                                 num_experts, num_experts_per_tok) for _ in range(num_blocks)]
+                                 num_experts, num_experts_per_tok, norm_topk_prob) for _ in range(num_blocks)]
     self.token_embd  = nn.Embedding(vocab_size, dim)
     self.output_norm = nn.RMSNorm(dim, norm_eps)
     self.output = nn.Linear(dim, vocab_size, bias=False)
@@ -221,7 +217,8 @@ class Transformer:
                         head_dim=kv.get(f'{arch}.attention.key_length', kv[f'{arch}.embedding_length'] // n_heads),
                         rope_theta=kv[f'{arch}.rope.freq_base'], max_context=max_context,
                         qk_norm=int(state_dict['blk.0.attn_q_norm.weight'].shape[0]) if 'blk.0.attn_q_norm.weight' in state_dict else 0,
-                        num_experts=kv.get(f'{arch}.expert_count', 0), num_experts_per_tok=kv.get(f'{arch}.expert_used_count', 0))
+                        num_experts=kv.get(f'{arch}.expert_count', 0), num_experts_per_tok=kv.get(f'{arch}.expert_used_count', 0),
+                        norm_topk_prob=True if arch=='qwen3moe' else False)
     nn.state.load_state_dict(model, state_dict, verbose=False, consume=True, realize=False)  # NOTE: rope_freqs.weight (32,) is unused
     # NOTE: without this contiguous, it unpacks the weights from the model every time. we shouldn't need this, but for now it's faster
     if realize:
@@ -239,10 +236,10 @@ class Transformer:
     t = Tensor(tokens + [0] * (self.max_context - len(tokens)), dtype="int32").reshape(1, self.max_context)
     # recompute start_pos from what's currently valid in the kv cache
     start_pos = self.get_start_pos(tokens)
-    out = None
+    out, prompt_len = None, len(tokens)
     while len(tokens) < self.max_context:
       sp, nt = v_start_pos.bind(start_pos), v_toks.bind(min(chunk_size, len(tokens) - start_pos))
-      out = self(t[:, sp:sp+nt] if out is None else out, sp).realize()
+      out = self(t[:, sp:sp+nt] if start_pos < prompt_len or out is None else out, sp).realize()
       start_pos += nt.val
       # chunked prefill: keep processing until all prompt tokens are consumed
       if start_pos < len(tokens): continue

tinygrad/codegen/__init__.py

@@ -19,7 +19,7 @@ from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_in
   ReduceContext, correct_load_store, pm_render, pm_add_loads
 from tinygrad.codegen.opt.postrange import apply_opts, pm_make_images
 from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse
-from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen, pm_mops, pm_syntactic_sugar
+from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen, pm_mops, pm_syntactic_sugar, pm_store_ranges
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize
 from tinygrad.renderer.amd.elf import do_assemble_amd
 
@@ -31,7 +31,7 @@ def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True) -
   if SPEC: type_verify(sink, kernel_spec)
 
   # preprocess
-  sink = graph_rewrite(sink, pm_mops+pm_syntactic_sugar, name="early movement ops", bottom_up=True)
+  sink = graph_rewrite(sink, pm_mops+pm_syntactic_sugar+pm_store_ranges, ctx=itertools.count(1000), name="early movement ops", bottom_up=True)
 
   # first we optimize
   if optimize:
@@ -48,7 +48,7 @@ def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True) -
     sink = graph_rewrite(sink, sym+pm_flatten_range, name="initial symbolic")
 
     # optimize (schedule) the AST
-    sink = graph_rewrite(sink, pm_simplify_ranges, name="simplify ranges")
+    sink = graph_rewrite(sink, pm_flatten_range+pm_simplify_ranges, ctx={}, name="simplify ranges")
 
     # do postrange optimization, BEAM or hand_coded_optimizations
     sink = apply_opts(sink, ren)

tinygrad/codegen/late/devectorizer.py

@@ -10,6 +10,7 @@ from tinygrad.renderer import Renderer
 
 # ***** image load valid simplification *****
 
+@functools.cache
 def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
   # can drop valid if idx is out of bound when valid is False
   drop_stmt = []
@@ -61,6 +62,18 @@ load_store_indexing = PatternMatcher([
 # ***** load/store grouping *****
 
 def expand_index(buf:UOp, vec:UOp):
+  # determine optimal image shapes
+  if IMAGE == 1 and isinstance(dt:=buf.dtype, ImageDType):
+    x, valid = vec.get_idx().gep(0), vec.get_valid().gep(0)
+    # search for dims that drop the most valid statements
+    best_drop, cands = -1, []
+    for ch, cw in ImageDType.valid_dims(dt):
+      if (dropped:=len(_drop_valid_stmts(valid, cidx:=uop_given_valid(valid, UOp.vectorize((x//4)%cw, x//(4*cw))), ch, cw))) > best_drop:
+        best_drop, cands = dropped, [(ch, cw, cidx)]
+      elif dropped == best_drop: cands.append((ch, cw, cidx))
+    # and tiebreak with indexing complexity (ie. number of nodes)
+    h, w, _ = cands[0] if len(cands) == 1 else min(cands, key=lambda cand: len(cand[2].gep(1).simplify().backward_slice))
+    buf = buf.replace(dtype=(dtypes.imageh if dt.itemsize == 2 else dtypes.imagef)((h, w, 4), w * 4 * dt.itemsize))
   if getenv("UNSAFE_DISABLE_MASK", 0): vec = vec.get_idx()
   # generate the individual indexes
   return UOp(Ops.VECTORIZE, buf.dtype, tuple(buf.index(vec.gep(i), ptr=True) for i in range(vec.dtype.count)))
@@ -183,31 +196,21 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
   if len(ret) <= 1: return None
   return UOp(Ops.VCAT, ls.dtype, tuple(ret)) if ls.op is Ops.LOAD else UOp.group(*ret)
 
-def _do_image_fixup(dt:ImageDType, idx:UOp) -> tuple[UOp, UOp, int, int]:
-  buf = idx.src[0]
-  x, valid = idx.src[1].get_idx(), idx.src[1].get_valid()
-  h, w = dt.shape[0], dt.shape[1]
-  if IMAGE == 1 and valid is not None:
-    h, w = max(ImageDType.valid_dims(dt), key=lambda hw:
-                # maximize number of valids removed
-               (len(_drop_valid_stmts(valid, idx:=uop_given_valid(valid, UOp.vectorize((x//4)%hw[1], x//(4*hw[1]))), *hw)),
-                # and minimize idx complexity (number of nodes)
-                -len(idx.simplify().backward_slice)))
-    buf = buf.replace(dtype=(dtypes.imageh if dt.itemsize == 2 else dtypes.imagef)((h, w, 4), w * 4 * dt.itemsize))
-  oidx = UOp(Ops.VECTORIZE, dtypes.index.vec(2), ((x // 4) % w, (x // (4*w))))
-  return x, idx.replace(src=(buf, oidx.valid(valid))), w, h
+def get_image_idx(idx:UOp, width:int):
+  oidx = UOp(Ops.VECTORIZE, dtypes.index.vec(2), (((x:=idx.src[1].get_idx()) // 4) % width, (x // (4*width))))
+  return idx.replace(src=(idx.src[0], oidx.valid(idx.src[1].get_valid())))
 
 def image_fixup(ls:UOp):
   # normal image load or store, with the CAST from expand_index
   if ls.src[0].op is Ops.CAST and isinstance(image_dtype:=ls.src[0].src[0].dtype, ImageDType):
     assert ls.src[0].dtype.count == 4, "image must be casted to 4"
-    _, idx, _, _ = _do_image_fixup(image_dtype, ls.src[0].src[0])
+    idx = get_image_idx(ls.src[0].src[0], image_dtype.shape[1])
     return ls.replace(src=(idx,)+ls.src[1:])
 
   # this is an unprocessed image without a cast, aka unfoldable image load. this doesn't work for stores
   if isinstance(image_dtype:=ls.src[0].dtype, ImageDType) and ls.src[0].src[1].get_idx().dtype != dtypes.index.vec(2):
     assert ls.op is Ops.LOAD, "if an image store isn't upcasted to 4, we can't store it"
-    x, idx, width, height = _do_image_fixup(image_dtype, ls.src[0])
+    x, idx = ls.src[0].src[1].get_idx(), get_image_idx(ls.src[0], image_dtype.shape[1])
     vec_load = ls.replace(dtype=ls.dtype.vec(4), src=(idx,)+ls.src[1:])
     # image pixels have 4 channels (.xyzw), select channel based on x % 4
     x_mod_4 = x % 4
@@ -248,32 +251,27 @@ def no_vectorized_alu(alu:UOp):
 def no_vectorized_buf(buf:UOp):
   return buf.replace(dtype=buf.ptrdtype.base.scalar().ptr(buf.ptrdtype.size*buf.ptrdtype.count, buf.ptrdtype.addrspace)).cast(buf.dtype)
 
-def no_vectorized_index(buf:UOp, cast:UOp, idx:UOp):
+def no_vectorized_index(buf:UOp, cast:UOp, idx:UOp, bcast:UOp|None=None):
   cnt = cast.dtype.count
-  assert idx.dtype.count == 1, f"idx dtype must be 1 {idx.dtype}"
-  return buf.broadcast(cnt).index(idx.broadcast(cnt)*cnt+UOp.const(dtypes.index.vec(cnt), tuple(range(cnt))), ptr=True)
-
-def no_vectorized_index_broadcast(buf:UOp, cast:UOp, bcast:UOp, idx:UOp):
-  cnt = cast.dtype.count
-  vcnt = cast.dtype.vcount
-  precnt = bcast.dtype.vcount
-  # TODO: I have no idea *why* this is. I just change things until the tests pass. No AI, old school.
-  if bcast.op is Ops.GEP:
-    gep_arg = tuple(flatten([range(precnt) for _ in range(vcnt)]))
-    sum_arg = tuple(flatten([[i+y for y in bcast.arg] for i in range(vcnt)]))
+  if bcast is not None and bcast.op is Ops.GEP:
+    # GEP selects specific lanes; bcast.arg[k] is the offset for lane k, iterate groups × selected lanes
+    pairs = [(k, g + bcast.arg[k]) for g, k in itertools.product(range(cast.dtype.vcount), range(len(bcast.arg)))]
+  elif bcast is not None:
+    # BROADCAST: cross product of components × lanes
+    pairs = [(j, c) for c, j in itertools.product(range(cnt), range(bcast.dtype.vcount))]
   else:
-    gep_arg = tuple(flatten([range(precnt) for _ in range(cnt)]))
-    sum_arg = tuple(flatten([[i]*precnt for i in range(cnt)]))
-  new_idx = idx.gep(gep_arg)*cnt + UOp.const(dtypes.index.vec(len(sum_arg)), sum_arg)
-  return buf.broadcast(cnt*precnt).index(new_idx, ptr=True)
+    # simple scalar index: one lane, all components
+    pairs = [(0, c) for c in range(cnt)]
+  idx_lanes, offsets = (tuple(x) for x in zip(*pairs))
+  return buf.broadcast(len(pairs)).index(idx.gep(idx_lanes)*cnt + UOp.const(dtypes.index.vec(len(pairs)), offsets), ptr=True)
 
 devectorize_buf_and_index = PatternMatcher([
   (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf"), no_vectorized_buf),
   (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG)).or_after(name="buf").cast(name="cast").index(UPat.var("idx")), no_vectorized_index),
   (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG)).or_after(name="buf").cast(name="cast").broadcast(name="bcast").index(UPat.var("idx")),
-   no_vectorized_index_broadcast),
+   no_vectorized_index),
   (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG)).or_after(name="buf").cast(name="cast").gep(name="bcast").index(UPat.var("idx")),
-   no_vectorized_index_broadcast),
+   no_vectorized_index),
 ])
 
 devectorize = PatternMatcher([

tinygrad/codegen/simplify.py

@@ -1,4 +1,5 @@
 import itertools
+from typing import Callable
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, graph_rewrite, _substitute, range_start
 from tinygrad.uop.symbolic import symbolic
 from tinygrad.helpers import partition
@@ -16,7 +17,7 @@ pm_flatten_range = PatternMatcher([
   (UPat((Ops.REDUCE, Ops.STORE, Ops.END), name="r"), flatten_range),
 ])
 
-def count_divmod(x:UOp) -> int: return len([u for u in x.toposort() if u.op in {Ops.IDIV, Ops.MOD}])
+def count_divmod(x:UOp) -> int: return sum(u.op in {Ops.IDIV, Ops.MOD} for u in x.backward_slice)
 def simplify_merge_adjacent(u:UOp) -> UOp|None:
   reduce_ranges = [x.ranges for x in u.backward_slice_with_self if x.op is Ops.REDUCE]
   # on END we only want to merge adjacent ranges, on REDUCE we want to try all combinations
@@ -36,22 +37,32 @@ def simplify_merge_adjacent(u:UOp) -> UOp|None:
           u = nidx
   return u
 
+def mark_gated(ctx, idx):
+  if idx.src[1].op is Ops.WHERE:
+    x, cond = idx.src[1].get_idx(), idx.src[1].get_valid()
+    # get all ranges r with guards "r < c" for some const c
+    guards = {r:c for v in cond.split_uop(Ops.AND) if v.op is Ops.CMPLT and (r:=v.src[0]).op is Ops.RANGE and (c:=v.src[1]).op is Ops.CONST}
+  else: x, guards = idx, {}
+  # ensure that we choose max(c_i) for all i where r < c_i
+  ctx |= {r:c for r,c in guards.items() if (r not in ctx or ctx[r].arg < c.arg)}
+  # but if a range is ever ungated, we cannot shrink it
+  ctx |= {r:r.src[0] for r in x.ranges if r not in guards}
+
 pm_simplify_ranges = PatternMatcher([
   (UPat((Ops.END, Ops.REDUCE), name="u"), simplify_merge_adjacent),
+  (UPat(Ops.INDEX, name="idx"), mark_gated),
+  # reduce ranges can't be shrunk
+  (UPat(Ops.REDUCE, name="red"), lambda ctx, red: ctx.update({r:r.src[0] for r in red.src[1:]})),
+  (UPat(Ops.SINK, name="x"), lambda ctx, x: do_substitute(ctx, x, lambda r,c: r.replace(src=(c,)))),
 ])
 
 def mark_range_mod(ctx:dict[UOp, UOp|None], r:UOp, c:UOp) -> None:
   if r not in ctx and r.src[0].op is Ops.CONST and r.src[0].divides(c.arg) is not None: ctx[r] = c
 
-def do_substitute(ctx:dict[UOp, UOp|None], x: UOp) -> UOp|None:
-  subs = {}
-  for k,v in ctx.items():
-    if v is not None:
-      subs[k] = k.replace(src=(k.src[0]//v,), arg=k.arg[0:-1]+(0,k.arg[-1]))*v + k.replace(src=(v,), arg=k.arg[0:-1]+(1,k.arg[-1]))
-  if not len(subs): return None
-  ret = x.substitute(subs).simplify()
+def do_substitute(ctx:dict, x: UOp, sub_fxn:Callable[[UOp, UOp], UOp]) -> UOp|None:
+  ret = x.substitute({k:sub_fxn(k,v) for k,v in ctx.items() if v is not None})
   ctx.clear()
-  return ret
+  return None if ret is x else ret.simplify()
 
 def dont_sub_ranges_for_image(ctx:dict[UOp, UOp|None], x:UOp) -> None:
   if isinstance(x.src[0].src[0].dtype, ImageDType):
@@ -60,7 +71,8 @@ def dont_sub_ranges_for_image(ctx:dict[UOp, UOp|None], x:UOp) -> None:
 pm_split_ranges = PatternMatcher([
   (UPat(Ops.RANGE, name="r")%UPat.cvar("c"), mark_range_mod),
   (UPat(Ops.STORE, name="x"), dont_sub_ranges_for_image),
-  (UPat(Ops.SINK, name="x"), do_substitute),
+  (UPat(Ops.SINK, name="x"), lambda ctx, x: do_substitute(ctx, x,
+    lambda k,v: k.replace(src=(k.src[0]//v,), arg=k.arg[0:-1]+(0,k.arg[-1]))*v + k.replace(src=(v,), arg=k.arg[0:-1]+(1,k.arg[-1])))),
 ])
 
 # **** reduce simplification ****

tinygrad/device.py

@@ -6,7 +6,7 @@ import importlib, inspect, functools, pathlib, os, platform, contextlib, sys, re
 from tinygrad.helpers import CI, OSX, LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored
 from tinygrad.helpers import Context, CCACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, dedup, ContextVar
 from tinygrad.helpers import unwrap_class_type, suppress_finalizing, select_first_inited, VIZ, CPU_LLVM, CPU_LVP, NV_PTX, CUDA_PTX, NV_NAK
-from tinygrad.helpers import EMULATED_DTYPES, TracingKey
+from tinygrad.helpers import EMULATED_DTYPES, NULL_IR3, NULL_QCOMCL, TracingKey
 from tinygrad.dtype import DType, ImageDType, PtrDType, dtypes, _to_np_dtype
 if TYPE_CHECKING: from tinygrad.renderer import Renderer
 
@@ -353,11 +353,12 @@ def is_dtype_supported(dtype:DType, device:str|None=None) -> bool:
     if device == "NV": return not CI and not NV_PTX and not NV_NAK
     if device in {"CPU"}: return not CI and platform.machine() in {"arm", "arm64", "aarch64", "x86_64", "amd64"} and not CPU_LVP
     return device in {"AMD", "CL", "PYTHON", "NULL"}
-  if dtype in dtypes.fp8s:
+  if dtype in dtypes.fp8_ocp:
     if device == "CUDA": return not CI and not CUDA_PTX
     if device == "NV": return not CI and not NV_PTX and not NV_NAK
-    if device == "AMD": return not CI and getattr(Device["AMD"], "target") in {(9,4,2), (9,5,0)}
+    if device == "AMD": return not CI and getattr(Device["AMD"], "target") == (9,5,0)
     return device in {"PYTHON", "NULL"}
+  if dtype in dtypes.fp8_fnuz: return device in {"PYTHON", "NULL"}
   if device == "WEBGPU": return dtype in [dtypes.bool, dtypes.char, dtypes.uchar, dtypes.short,
                                           dtypes.ushort, dtypes.float, dtypes.int32, dtypes.uint32, dtypes.half]
   # for CI GPU and OSX, cl_khr_fp16 isn't supported
@@ -371,7 +372,7 @@ def is_dtype_supported(dtype:DType, device:str|None=None) -> bool:
     if device in ["CUDA", "NV"]: return not CI
     if device == "CPU" and CPU_LLVM: return OSX
     if device == "PYTHON": return sys.version_info >= (3, 12)
-  if dtype == dtypes.float64: return (device not in {"METAL", "QCOM"} and not (OSX and device == "CL") and not getenv("NULL_IR3")
+  if dtype == dtypes.float64: return (device not in {"METAL", "QCOM"} and not (OSX and device == "CL") and not NULL_IR3 and not NULL_QCOMCL
                                       and dtypes.long not in EMULATED_DTYPES.tolist(dtypes))
   return True
 

tinygrad/dtype.py

@@ -30,6 +30,7 @@ class InvalidType:
   def __hash__(self): return id(self)
   def __repr__(self): return "Invalid"
   def __reduce__(self): return (InvalidType, ())  # unpickle returns the singleton
+  def __format__(self, spec): return "Invalid"
 
 Invalid = InvalidType()
 
@@ -78,10 +79,14 @@ class DType(metaclass=DTypeMetaClass):
     return PtrDType(self.priority, self.bitsize, self.name, self.fmt, self.count, None, self, addrspace, 1, size)
   def scalar(self) -> DType: return self._scalar if self._scalar is not None else self
   def nbytes(self) -> int: raise RuntimeError("only ptr types have nbytes")
-  @property
-  def min(self): return dtypes.min(self)
-  @property
-  def max(self): return dtypes.max(self)
+  @functools.cached_property
+  def min(self):
+    if dtypes.is_int(self): return 0 if dtypes.is_unsigned(self) else -2**(self.scalar().bitsize-1)
+    return -float("inf") if dtypes.is_float(self) else False
+  @functools.cached_property
+  def max(self):
+    if dtypes.is_int(self): return 2**(self.scalar().bitsize)-1+self.min
+    return float("inf") if dtypes.is_float(self) else True
 
 @dataclass(frozen=True, eq=False)
 class PtrDType(DType):
@@ -171,21 +176,11 @@ class dtypes:
     # int is the default. wrap floats in ConstFloat to distinguish -0.0 from 0.0 in cache
     return ConstFloat(float(val)) if dtypes.is_float(dtype) else bool(val) if dtypes.is_bool(dtype) else int(val)
   @staticmethod
-  @functools.cache
-  def min(dtype:DType):
-    if dtypes.is_int(dtype): return 0 if dtypes.is_unsigned(dtype) else -2**(dtype.scalar().bitsize-1)
-    return -float("inf") if dtypes.is_float(dtype) else False
-  @staticmethod
-  @functools.cache
-  def max(dtype:DType):
-    if dtypes.is_int(dtype): return 2**(dtype.scalar().bitsize)-1+dtypes.min(dtype)
-    return float("inf") if dtypes.is_float(dtype) else True
-  @staticmethod
   def finfo(dtype:DType) -> tuple[int, int]:
     """(exponent, mantissa)"""
     if not dtypes.is_float(dtype): raise ValueError(f"{dtype} is not a floating point type")
     return {dtypes.float16: (5, 10), dtypes.bfloat16: (8, 7), dtypes.float32: (8, 23), dtypes.float64: (11, 52),
-            dtypes.fp8e5m2: (5, 2), dtypes.fp8e4m3: (4, 3)}[dtype]
+            dtypes.fp8e4m3: (4, 3), dtypes.fp8e5m2: (5, 2), dtypes.fp8e4m3fnuz: (4, 3), dtypes.fp8e5m2fnuz: (5, 2)}[dtype]
   void: Final[DType] = DType.new(-1, 0, "void", None)
   index: Final[DType] = DType.new(-1, 800, "index", None)
   bool: Final[DType] = DType.new(0, 1, "bool", '?')
@@ -201,6 +196,8 @@ class dtypes:
   _uint256: Final[DType] = DType.new(8, 256, "uint256", None)
   fp8e4m3: Final[DType] = DType.new(9, 8, "float8_e4m3", None)
   fp8e5m2: Final[DType] = DType.new(10, 8, "float8_e5m2", None)
+  fp8e4m3fnuz: Final[DType] = DType.new(9, 8, "float8_e4m3fnuz", None)
+  fp8e5m2fnuz: Final[DType] = DType.new(10, 8, "float8_e5m2fnuz", None)
   float16: Final[DType] = DType.new(11, 16, "half", 'e')
   # bfloat16 has higher priority than float16, so least_upper_dtype(dtypes.int64, dtypes.uint64) = dtypes.float16
   bfloat16: Final[DType] = DType.new(12, 16, "__bf16", None)
@@ -221,7 +218,9 @@ class dtypes:
   default_float: ClassVar[DType] = float32
   default_int: ClassVar[DType] = int32
 
-  fp8s = (fp8e4m3, fp8e5m2)
+  fp8_ocp = (fp8e4m3, fp8e5m2)
+  fp8_fnuz = (fp8e4m3fnuz, fp8e5m2fnuz)
+  fp8s = fp8_ocp + fp8_fnuz
   floats = fp8s + (float16, bfloat16, float32, float64)
   int8s = (uint8, int8)
   int16s = (uint16, int16)
@@ -243,8 +242,9 @@ def to_dtype(dtype:DTypeLike) -> DType: return dtype if isinstance(dtype, DType)
 # we don't support weak type and complex type
 promo_lattice = { dtypes.bool: [dtypes.int8, dtypes.uint8], dtypes.int8: [dtypes.int16], dtypes.int16: [dtypes.int32], dtypes.int32: [dtypes.int64],
   dtypes.int64: [dtypes.uint64], dtypes.uint8: [dtypes.int16, dtypes.uint16], dtypes.uint16: [dtypes.int32, dtypes.uint32],
-  dtypes.uint32: [dtypes.int64, dtypes.uint64], dtypes.uint64: [dtypes.fp8e4m3, dtypes.fp8e5m2],
-  dtypes.fp8e5m2: [dtypes.float16, dtypes.bfloat16], dtypes.fp8e4m3: [dtypes.float16, dtypes.bfloat16],
+  dtypes.uint32: [dtypes.int64, dtypes.uint64], dtypes.uint64: [dtypes.fp8e4m3, dtypes.fp8e5m2, dtypes.fp8e4m3fnuz, dtypes.fp8e5m2fnuz],
+  dtypes.fp8e4m3: [dtypes.float16, dtypes.bfloat16], dtypes.fp8e5m2: [dtypes.float16, dtypes.bfloat16],
+  dtypes.fp8e4m3fnuz: [dtypes.float16, dtypes.bfloat16], dtypes.fp8e5m2fnuz: [dtypes.float16, dtypes.bfloat16],
   dtypes.float16: [dtypes.float32], dtypes.bfloat16: [dtypes.float32], dtypes.float32: [dtypes.float64], }
 
 @functools.cache
@@ -299,10 +299,14 @@ def float_to_bf16(x):
 _fp8_cfg = {
   dtypes.fp8e4m3: (7, 4, 0x7, 0x3F50000000000000, 0x407D000000000000, 0x7E, 0x3F90000000000000),
   dtypes.fp8e5m2: (15, 3, 0x3, 0x3EE0000000000000, 0x40EE000000000000-1, 0x7B, 0x3F10000000000000),
+  dtypes.fp8e4m3fnuz: (8, 4, 0x7, 0x3F40000000000000, 0x406F000000000000-1, 0x7F, 0x3F80000000000000),
+  dtypes.fp8e5m2fnuz: (16, 3, 0x3, 0x3ED0000000000000, 0x40EE000000000000-1, 0x7F, 0x3F00000000000000),
 }
 
 def float_to_fp8(x: float, dtype: DType) -> int:
   assert dtype in dtypes.fp8s, "Only for fp8s"
+  if dtype in dtypes.fp8_fnuz and not math.isfinite(x): return 0x80
+  if dtype in dtypes.fp8_fnuz and x == 0.0: return 0x00
   # e4m3 don't support inf, return 0x7f(+NaN) and 0xff(-NaN) to match jax
   # NaN is unordered, can't compare with zero, use math.copysign to get sign
   if dtype == dtypes.fp8e4m3 and not math.isfinite(x): return 0x7f if math.copysign(1, x) > 0 else 0xff
@@ -322,16 +326,17 @@ def float_to_fp8(x: float, dtype: DType) -> int:
     res, half = mantissa >> shift, half_ulp << shift
     round_bits = (xbits | (1 << 52)) & ((half << 1) - 1)
     if round_bits > half or (round_bits == half and res & 1): res += 1
-  return int(res | sign)
+  return 0 if dtype in dtypes.fp8_fnuz and res == 0 else int(res | sign)  # fnuz has no negative zero
 
 def fp8_to_float(x: int, dtype: DType) -> float:
   assert dtype in dtypes.fp8s, "Only for fp8s"
+  if dtype in dtypes.fp8_fnuz and x == 0x80: return math.nan
   if (x & 0x7F) == 0: return -0.0 if x & 0x80 else 0.0
   bias, sig_bits, *_ = _fp8_cfg[dtype]
   mant_bits, exp_bits = sig_bits - 1, 8 - sig_bits
   exp_max, mant_max = (1 << exp_bits) - 1, (1 << mant_bits) - 1
   sign, exp, mantissa = (x >> 7) & 1, (x >> mant_bits) & exp_max, x & mant_max
-  if exp == exp_max:
+  if dtype not in dtypes.fp8_fnuz and exp == exp_max:
     if dtype == dtypes.fp8e5m2: return math.copysign(math.nan if mantissa else math.inf, -1 if sign else 1)
     if mantissa == mant_max: return math.nan
   val = (mantissa / (mant_max + 1)) * 2 ** (1 - bias) if exp == 0 else (1 + mantissa / (mant_max + 1)) * 2 ** (exp - bias)

tinygrad/engine/allocations.py

@@ -1,7 +1,7 @@
 from dataclasses import dataclass, field
 from tinygrad.uop.ops import UOp, UPat, PatternMatcher, Ops, GroupOp, graph_rewrite, track_rewrites
-from tinygrad.dtype import dtypes, ImageDType
-from tinygrad.helpers import prod, DEBUG, VIZ, pluralize
+from tinygrad.dtype import ImageDType
+from tinygrad.helpers import prod, DEBUG, VIZ, pluralize, all_int
 
 @dataclass
 class AllocCtx:
@@ -29,14 +29,15 @@ def apply_after(ctx:AllocCtx, u:UOp):
   while base.op is Ops.AFTER: base = base.src[0]
   ctx.buffer_map[u] = base
 
-# CONTIGUOUS and ASSIGN + parents are the only nodes that get updated
+# CONTIGUOUS and AFTER+STORE + parents are the only nodes that get updated
 add_tags = PatternMatcher([
   (UPat(Ops.COPY, name="u"), disk_copy_is_buffer),
-  # no tag on copies that are assigned
-  (UPat(Ops.ASSIGN, src=(UPat(), UPat(Ops.COPY, name="c")), name="a"),
-   lambda a,c: a.replace(src=(a.src[0], c.rtag(())), tag=a.tag+c.tag) if a.tag and c.tag else None),
+  # no tag on copies that are assigned via STORE+AFTER — merge COPY tag into AFTER
+  (UPat(Ops.AFTER, src=(UPat(), UPat(Ops.STORE, src=(UPat(name="dest"), UPat(Ops.COPY, name="c")))), name="a"),
+   lambda a,c,dest: a.replace(src=(a.src[0], a.src[1].replace(src=(dest, c.rtag(())))), tag=a.tag+c.tag) if a.tag and c.tag else None),
+  (UPat(Ops.AFTER, src=(UPat(), UPat(Ops.STORE)), name="x"), tag_uop),
   (UPat(Ops.AFTER, name="u"), apply_after),
-  (UPat({Ops.CONTIGUOUS, Ops.ASSIGN}, name="x"), tag_uop),
+  (UPat(Ops.CONTIGUOUS, name="x"), tag_uop),
   (UPat(GroupOp.All, name="x"), lambda ctx,x: tag_uop(ctx,x) if x in ctx.bases else None),
 ])
 
@@ -46,34 +47,33 @@ def _buffer_like(u:UOp) -> UOp:
     if prod(dtype.shape) != prod(u.max_shard_shape) or ([x for x in u.max_shard_shape if x != 1] or [1])[-1] % 4 != 0:
       if DEBUG >= 1: print(f"demoting Image {dtype} with shape {u.max_shard_shape}")
       dtype = dtype.base
-  buffer = UOp.new_buffer(u.device, u.shard_size, dtype).reshape(u.max_shard_shape)
+  buffer = UOp.new_buffer(u.device, u.shard_size, dtype).reshape(u.max_shard_shape).shrink_to(u.shard_shape)
   if isinstance(u.device, tuple) and u.axis is not None: buffer = buffer.multi(u.axis)
   return buffer
 
-def replace_contig_with_assign(u:UOp):
+def replace_contig_with_store_after(u:UOp):
   # can't allocate a buffer without a device (e.g., inside a CALL function body with only PARAMs)
   if u._device is None: return None
   # if size is 0, remove the contig
   if u.size == 0: return u.src[0]
   # no real contig for DISK/TINYFS tensors, they are left alone
   if isinstance(u._device, str) and u._device.startswith(("DISK", "TINYFS")): return u.rtag(None)
-  return _buffer_like(u).assign(u.src[0]).rtag(u.tag)
+  buf = _buffer_like(u)
+  return buf.after(buf.store(u.src[0])).rtag(u.tag)
 
-def replace_assign_with_contig(u:UOp):
+def replace_store_after_with_contig(u:UOp, src:UOp):
   assigned_to = u
-  while assigned_to.op in {Ops.ASSIGN, Ops.BITCAST, Ops.AFTER}: assigned_to = assigned_to.src[0].base
-  if assigned_to.op is not Ops.BUFFER:
-    return u.src[1].contiguous(tag=u.tag)
+  while assigned_to.op in {Ops.BITCAST, Ops.AFTER}: assigned_to = assigned_to.src[0].base
+  if assigned_to.op is not Ops.BUFFER: return src.contiguous(tag=u.tag)
 
-def contiguous_mops_to_view(c:UOp):
+def contiguous_mops_to_view(c:UOp, src:UOp):
   """CONTIGUOUS(MOPS(BUFFER)) → CONTIGUOUS(BUFFER_VIEW) when movement ops collapse to a contiguous range."""
-  src = c.src[0]
   buf = src.base
   if buf.op not in {Ops.BUFFER, Ops.BUFFER_VIEW}: return None
   if src.op is Ops.RESHAPE and src.src[0].op in {Ops.BUFFER, Ops.BUFFER_VIEW}: return None
 
   # no symbolic shape
-  if not all(isinstance(x, int) for x in c.shape): return None
+  if not all_int(c.shape): return None
 
   # check if view is supported
   if not isinstance(c.device, str): return None
@@ -81,8 +81,7 @@ def contiguous_mops_to_view(c:UOp):
   if not hasattr(Device[c.device].allocator, "_offset"): return None
 
   # see if this can be a view
-  offset = src.contiguous_view_offset()
-  if offset is None: return None
+  if (offset := src.contiguous_view_offset()) is None: return None
 
   # merge BUFFER_VIEWs
   if buf.op is Ops.BUFFER_VIEW: offset, buf = offset + buf.arg[1], buf.src[0]
@@ -93,31 +92,47 @@ def contiguous_mops_to_view(c:UOp):
 def transform_precompiled_call(c:UOp) -> UOp|None:
   if not c.arg.precompile: return None
   if c.src[0].op is Ops.SINK: return None
-  out = _buffer_like(c)
+  assert c.src[0].op is Ops.TUPLE, f"expected TUPLE body for precompiled call, got {c.src[0].op}"
   input_buffers = tuple(x.contiguous() if x.op not in {Ops.AFTER, Ops.BIND} else x for x in c.src[1:])
-  fxn = out.param_like(len(c.src)-1).assign(c.src[0]).sink()
-  ret = out.after(c.replace(src=(fxn, *input_buffers, out), dtype=dtypes.void, tag=None))
+
+  # add the outputs to the call
+  srcs = c.src[0].src
+  resolved = [c.gettuple(i) for i in range(len(srcs))]
+  outs = tuple(_buffer_like(r) for r in resolved)
+  targets = [o.param_like(len(c.src)-1+i).shrink_to(s.shape) for i,(o,s) in enumerate(zip(outs, srcs))]
+  fxn = UOp.sink(*[t.after(t.store(s)) for t,s in zip(targets, srcs)])
+
+  # create the new thing for the big graph
+  new_call = c.replace(src=(fxn, *input_buffers, *outs), tag=None)
+  rets = tuple(o.after(new_call) for o in outs)
+
   # if the CALL has symbolic shapes, shrink the max-sized output to the actual symbolic shape
-  if any(isinstance(s, UOp) for s in c.shape): ret = ret.shrink(tuple((0, s) for s in c.shape))
-  return ret
+  # NOTE: must use resolved shapes from the CALL (which substitutes PARAMs with external args), not raw body shapes
+  rets = tuple(r.shrink_to(rs.shape) for r,rs in zip(rets, resolved))
+
+  return UOp.maketuple(*rets)
 
 # NOTE: adding rules to here is bad. these all need to run before the schedule cache
 pm_early_transform_tensor_graph = PatternMatcher([
   # transform precompiled CALLs
   (UPat(Ops.CALL, name="c"), transform_precompiled_call),
 
+  # resolve TUPLE+GETTUPLE (for precompiled calls)
+  (UPat(Ops.GETTUPLE, src=(UPat(Ops.TUPLE, name="t"),), name="g"), lambda g,t: t.src[g.arg]),
+
   # CONTIGUOUS(MOPS(BUFFER/BUFFER_VIEW)) → CONTIGUOUS(BUFFER_VIEW) when movement ops collapse to contiguous range
-  (UPat(Ops.CONTIGUOUS, src=(UPat(GroupOp.Movement),), name="c"), contiguous_mops_to_view),
+  (UPat(Ops.CONTIGUOUS, src=(UPat(GroupOp.Movement, name="src"),), name="c"), contiguous_mops_to_view),
 
   # add CONTIGUOUS to tagged UOps
-  (UPat(GroupOp.All-{Ops.CONTIGUOUS, Ops.ASSIGN}, name="x"), lambda x: x.rtag(None).contiguous(tag=x.tag) if x.tag else x.replace(tag=None)),
-  # remove extra CONTIGUOUS on ASSIGN (only when assign target is contiguous)
-  (UPat(Ops.CONTIGUOUS, src=(UPat(Ops.ASSIGN, name="a"),), name="c"),
+  (UPat(GroupOp.All-{Ops.CONTIGUOUS, Ops.AFTER, Ops.STORE}, name="x"),
+   lambda x: x.rtag(None).contiguous(tag=x.tag) if x.tag else x.replace(tag=None)),
+  # remove extra CONTIGUOUS on AFTER (only when target is contiguous)
+  (UPat(Ops.CONTIGUOUS, src=(UPat(Ops.AFTER, name="a"),), name="c"),
    lambda a,c: a.replace(tag=(a.tag or ())+(c.tag or ())) if a.src[0].has_buffer_identity() else None),
-  # replace ASSIGN with CONTIGUOUS
-  (UPat(Ops.ASSIGN, name="u"), replace_assign_with_contig),
-  # replace CONTIGUOUS with ASSIGNs
-  (UPat(Ops.CONTIGUOUS, name="u"), replace_contig_with_assign),
+  # replace AFTER+STORE with CONTIGUOUS when target is not a buffer
+  (UPat(Ops.AFTER, src=(UPat(), UPat(Ops.STORE, src=(UPat(), UPat(name="src")))), name="u"), replace_store_after_with_contig),
+  # replace CONTIGUOUS with STORE+AFTER
+  (UPat(Ops.CONTIGUOUS, name="u"), replace_contig_with_store_after),
   # remove DETACH/CONTIGUOUS_BACKWARD (allows more contiguous removal)
   (UPat((Ops.DETACH, Ops.CONTIGUOUS_BACKWARD), name="x"), lambda x: x.src[0]),
 ])
@@ -128,9 +143,9 @@ def untag_and_append(ctx:AllocCtx, x:UOp):
   for t in x.tag:
     original_uop: UOp = ctx.uop_list[t]
     replace_uop = ret
-    while replace_uop.op is Ops.ASSIGN: replace_uop = replace_uop.src[0]
+    while replace_uop.op is Ops.AFTER: replace_uop = replace_uop.src[0]
     ctx.buffer_map[original_uop] = replace_uop.shrink_to(original_uop.shape)
-  ctx.assigns.append(ret)
+  if ret.op is not Ops.AFTER: ctx.assigns.append(ret)  # AFTER gets appended by append_after
   return ret
 
 def append_after(ctx:AllocCtx, x:UOp):
@@ -142,7 +157,7 @@ def replace_input_buffer(ctx:AllocCtx, b:UOp):
                    b._min_max if b.op is Ops.BIND else None, b.src[0].arg[0] if b.op is Ops.BIND else None)
 
 pm_finalize_call = PatternMatcher([
-  (UPat(Ops.ASSIGN, name="x"), untag_and_append),
+  (UPat(Ops.AFTER, name="x"), untag_and_append),
   (UPat(Ops.AFTER, name="x"), append_after),
   (UPat(Ops.COPY, name="x"), lambda ctx,x: append_after(ctx,x) if isinstance(x.device, str) and x.device.startswith(("DISK", "TINYFS")) else None),
   # remove unique from const. TODO: this is copied in function.py

tinygrad/engine/jit.py

@@ -114,9 +114,9 @@ class GraphRunner(Runner):
           assert ji.prg.p.local_size is not None
           self.launch_dims_base[j] = (tuple(ji.prg.p.global_size), tuple(ji.prg.p.local_size))
 
-    # used in MultiGraphRunner. the ints are id() of _bufs
-    self.w_dependency_map: dict[int, Any] = {}
-    self.r_dependency_map: dict[int, list[Any]] = collections.defaultdict(list)
+    # used in MultiGraphRunner. tracks (offset, end, dep) ranges per base buffer id to handle suballocated buffers correctly.
+    self.w_dependency_map: dict[int, list[tuple[int, int, Any]]] = collections.defaultdict(list)
+    self.r_dependency_map: dict[int, list[tuple[int, int, Any]]] = collections.defaultdict(list)
 
     assert jit_cache[0].prg is not None
     super().__init__(colored(f"<batched {len(jit_cache)}>", "cyan"), jit_cache[0].prg.device.split(":")[0], estimates.simplify())
@@ -132,19 +132,22 @@ class GraphRunner(Runner):
       yield j, (dims[gl] if gl is not None else self.launch_dims_base[j][0]), (dims[lc] if lc is not None else self.launch_dims_base[j][1])
 
   def _access_resources(self, bufs:list[Buffer], write:list[int], new_dependency:Any):
-    # To synchronize access to resources, we monitor the necessary prerequisites for accessing each resource,
-    # whether for write or read operations. A resource can be accessed by either a single writer or multiple readers.
     wait_nodes = []
-
     for i,buf in enumerate(bufs):
-      if id(buf.base._buf) in self.w_dependency_map: wait_nodes.append(self.w_dependency_map[id(buf.base._buf)])
+      key, s, e = id(buf.base._buf), buf.offset, buf.offset + buf.nbytes
+      wait_nodes += [dep for st,en,dep in self.w_dependency_map[key] if st < e and s < en]
+      if i in write: wait_nodes += [dep for st,en,dep in self.r_dependency_map[key] if st < e and s < en]
+    for i,buf in enumerate(bufs):
+      key, s, e = id(buf.base._buf), buf.offset, buf.offset + buf.nbytes
       if i in write:
-        if id(buf.base._buf) in self.r_dependency_map: wait_nodes.extend(self.r_dependency_map.pop(id(buf.base._buf)))
-
-    for i,buf in enumerate(bufs):
-      if i in write: self.w_dependency_map[id(buf.base._buf)] = new_dependency
-      else: self.r_dependency_map[id(buf.base._buf)].append(new_dependency)
-
+        for dmap in [self.w_dependency_map, self.r_dependency_map]:
+          kept = []
+          for st,en,dep in dmap[key]:
+            if st < min(s, en): kept.append((st, min(s, en), dep))
+            if max(e, st) < en: kept.append((max(e, st), en, dep))
+          dmap[key] = kept
+        self.w_dependency_map[key].append((s, e, new_dependency))
+      else: self.r_dependency_map[key].append((s, e, new_dependency))
     return list({id(x):x for x in wait_nodes}.values())
 
   @staticmethod
@@ -357,9 +360,8 @@ class TinyJit(Generic[ReturnType]):
         jit_cache = pruned
 
       # memory planning (optional)
-      # Exclude buffers involved in transfer ops to preserve parallelism.
-      noopt_buffers = {b for ji in jit_cache if isinstance(ji.prg, (BufferXfer, BufferCopy, EncDec)) for b in ji.bufs}
-      assigned = _internal_memory_planner([cast(list[Buffer], item.bufs) for item in jit_cache], noopt_buffers, debug_prefix="JIT ")
+      copies = [(cast(Buffer,ji.bufs[0]),cast(Buffer,ji.bufs[1])) for ji in jit_cache if isinstance(ji.prg, (BufferXfer, BufferCopy, EncDec))]
+      assigned = _internal_memory_planner([cast(list[Buffer], item.bufs) for item in jit_cache], copies, debug_prefix="JIT ")
       jit_cache = [replace(item, bufs=[assigned.get(b,b).ensure_allocated() for b in item.bufs if b is not None]) for item in jit_cache]
 
       input_replace = get_input_replace(jit_cache, input_buffers)

tinygrad/engine/memory.py

@@ -7,43 +7,50 @@ from tinygrad.uop.ops import Ops
 from tinygrad.dtype import dtypes, ImageDType
 from tinygrad.runtime.support.memory import TLSFAllocator
 
+LaneKey = tuple[str, int]
+
 # **************** memory planning ****************
 
-def _internal_memory_planner(buffers:list[list[Buffer]], noopt_buffers=None, ignore_checks=False, debug_prefix="") -> dict[Buffer, Buffer]:
+def _internal_memory_planner(buffers:list[list[Buffer]], copies:list[tuple[Buffer, Buffer]]|None=None,
+                             ignore_checks=False, debug_prefix="") -> dict[Buffer, Buffer]:
   if NO_MEMORY_PLANNER: return {}
   first_appearance, last_appearance, buf_to_opt = {}, {}, set()
   for i,u in enumerate(buffers):
     for buf in u:
-      should_skip = buf.is_allocated() or buf.base.is_allocated() or buf.uop_refcount > 0 or (noopt_buffers is not None and buf.base in noopt_buffers)
-      if not ignore_checks and should_skip: continue
+      if not ignore_checks and (buf.is_allocated() or buf.base.is_allocated() or buf.uop_refcount > 0): continue
       if buf.base not in first_appearance: first_appearance[buf.base] = i
       last_appearance[buf.base] = i
       buf_to_opt.add(buf)
 
+  # Separate copy and compute buffers into different lanes and defer cross-queue frees to avoid introducing dependencies (copy->compute->copy)
+  copy_dsts, copy_srcs = ({dst.base for dst,_ in copies}, {src.base for _,src in copies}) if copies else (set(), set())
+  def _key(buf) -> LaneKey: return (buf.device, 1 if buf in copy_dsts or buf in copy_srcs else 0)
+  buf_hold = {buf: last_appearance[buf] - first_appearance[buf] + 1 for buf in first_appearance if buf in copy_dsts or buf in copy_srcs}
+
   # Sort buffer operations in timeline order. Two events: buffer is allocated or buffer is freed.
   buffer_requests = sorted([((first_appearance[buf], True), buf) for buf in first_appearance.keys()] + \
-                           [((last_appearance[buf] + 1, False), buf) for buf in first_appearance.keys()], key=lambda x: x[0])
-  total_memory = sum(round_up(buf.nbytes, min_block_size:=0x1000) for buf in first_appearance.keys()) * 2 # *2 for fragmentation (which is about 15%)
+    [((last_appearance[buf] + 1 + buf_hold.get(buf, 0), False), buf) for buf in first_appearance.keys()], key=lambda x: x[0])
+  total_memory = sum(round_up(buf.nbytes, BLK:=0x1000) for buf in first_appearance.keys()) * 2 # *2 for fragmentation (which is about 15%)
 
   # Try to suballocate from a shared buffer managed by global_planner using TLSFAllocator.
   # Also track buffer replacements for buffers that do not support suballocation.
   buffer_replace:dict[Buffer, tuple[Buffer|None, int|None]] = {}
   reuse_buffers:dict[tuple, list[Buffer]] = defaultdict(list)
-  global_planner:dict[str, tuple[int, TLSFAllocator]] = defaultdict(lambda: (0, TLSFAllocator(total_memory, block_size=min_block_size, lv2_cnt=32)))
+  global_planner:dict[LaneKey, tuple[int, TLSFAllocator]] = defaultdict(lambda: (0, TLSFAllocator(total_memory, block_size=BLK, lv2_cnt=32)))
   for (_, is_open_ev), buf in buffer_requests:
     # Check if suballocation is possible for the given buffer and device.
     if hasattr(Device[buf.device].allocator, "_offset") and not isinstance(buf.dtype, ImageDType):
-      if is_open_ev: buffer_replace[buf] = (None, global_planner[buf.device][1].alloc(round_up(buf.nbytes, 0x1000)))
-      else: global_planner[buf.device][1].free(cast(int, buffer_replace[buf][1]))
-      global_planner[buf.device] = (max(global_planner[buf.device][0], buffer_replace[buf][1] + buf.nbytes), global_planner[buf.device][1])
+      if is_open_ev: buffer_replace[buf] = (None, global_planner[_key(buf)][1].alloc(round_up(buf.nbytes, BLK)))
+      else: global_planner[_key(buf)][1].free(cast(int, buffer_replace[buf][1]))
+      global_planner[_key(buf)] = (max(global_planner[_key(buf)][0], buffer_replace[buf][1] + buf.nbytes), global_planner[_key(buf)][1])
     else:
-      key = (buf.device, buf.dtype, buf.options, buf.nbytes)
+      key = (_key(buf), buf.dtype, buf.options, buf.nbytes)
       if is_open_ev: buffer_replace[buf] = (reuse_buffers[key].pop(), None) if key in reuse_buffers and len(reuse_buffers[key]) > 0 else (buf, None)
       else: reuse_buffers[key].append(cast(Buffer, buffer_replace[buf][0]))
 
   # Allocate global buffers based on the memory planner.
-  global_buffers = {dev: Buffer(dev, round_up(sz, 0x1000), dtypes.int8) for dev, (sz, _) in global_planner.items()}
-  buffer_resolve:dict[Buffer, tuple[Buffer, int|None]] = {buf: (base or global_buffers[buf.device], off) for buf,(base,off) in buffer_replace.items()}
+  global_buffers = {key: Buffer(key[0], round_up(sz, BLK), dtypes.int8) for key, (sz, _) in global_planner.items()}
+  buffer_resolve:dict[Buffer, tuple[Buffer, int|None]] = {buf: (base or global_buffers[_key(buf)], off) for buf,(base,off) in buffer_replace.items()}
 
   # Assign buffers. First, assign full buffers (not sub-buffers).
   assigned:dict[Buffer, Buffer] = {}
@@ -66,5 +73,5 @@ def _internal_memory_planner(buffers:list[list[Buffer]], noopt_buffers=None, ign
 def memory_planner(schedule:list[ExecItem]) -> list[ExecItem]:
   # Exclude buffers involved in load ops (e.g transfers) to preserve parallelism in graphs.
   assigned = _internal_memory_planner([[b for b in si.bufs if b is not None] for si in schedule],
-                                      noopt_buffers={b for si in schedule if si.ast.op is not Ops.SINK for b in si.bufs if b is not None})
+                                      copies=[(cast(Buffer,si.bufs[0]),cast(Buffer,si.bufs[1])) for si in schedule if si.ast.op is Ops.COPY])
   return [ExecItem(si.ast, [assigned.get(x, x) if x is not None else None for x in si.bufs], si.metadata, si.fixedvars) for si in schedule]

tinygrad/engine/realize.py

@@ -93,8 +93,8 @@ class BufferXfer(BufferCopy):
   def copy(self, dest, src): dest.allocator._transfer(dest._buf, src._buf, dest.nbytes, src_dev=src.allocator.dev, dest_dev=dest.allocator.dev)
 
 class EncDec(Runner):
-  def __init__(self, encdec:UOp, total_sz:int, device:str):
-    self.shape, self.pos_var = encdec.arg[0], encdec.variables()[0].expr
+  def __init__(self, cf:UOp, total_sz:int, device:str):
+    self.shape, self.pos_var = tuple(s.arg for s in cf.src if s.op is Ops.CONST), cf.variables()[0].expr
     name = f"enc/dec {total_sz/1e6:7.2f}M, HEVC" if total_sz >= 1e6 else f"enc/dec {total_sz:8d}, HEVC"
     super().__init__(colored(name, "yellow"), device, Estimates(lds=total_sz, mem=total_sz))
   def __call__(self, rawbufs:list[Buffer], var_vals:dict[str, int], wait=False):
@@ -130,7 +130,7 @@ si_lowerer = PatternMatcher([
   (UPat(Ops.COPY, name="copy"), lambda ctx,copy: (BufferXfer(ctx[0].nbytes, ctx[0].device, ctx[1].device) \
       if hasattr(alc:=Device[ctx[0].device].allocator, '_transfer') and alc.supports_transfer and all_same([x.device.split(":")[0] for x in ctx]) \
       else BufferCopy(ctx[0].nbytes, ctx[0].device, ctx[1].device))),
-  (UPat(Ops.ENCDEC, name="encdec"), lambda ctx,encdec: EncDec(encdec, ctx[0].nbytes, ctx[1].device)),
+  (UPat(Ops.CUSTOM_FUNCTION, arg="encdec", name="cf"), lambda ctx,cf: EncDec(cf, ctx[0].nbytes, ctx[0].device)),
 ])
 
 @dataclass

tinygrad/engine/schedule.py

@@ -22,6 +22,7 @@ def create_schedule(sched_sink:UOp) -> UOp:
     for u in sched_sink.toposort(gate_kernel_sink):
       if u.op is not Ops.AFTER: continue
       k = u.src[1]
+      if k.op is Ops.STORE: continue  # skip unprocessed STORE+AFTER inside precompiled CALL bodies
       assert k.op in {Ops.CALL, Ops.END}, f"AFTER src[1] should be CALL or END, not {k.op}"
       in_degree.setdefault(k, 0)
       if k.op is Ops.END: assert k.src[0].op is Ops.CALL, f"END src[0] should be KERNEL, not {k.src[0].op}"
@@ -71,7 +72,7 @@ def linear_to_schedule(linear:UOp) -> list[ExecItem]:
       base = buf_uops[1].buffer
       assert isinstance(base, Buffer), "base can't be MultiBuffer"
       buffers[buf_uops[0]] = base.view(buf_uops[0].arg, ast.dtype, ast.arg[1]*base.dtype.itemsize)
-    ubufs = [b.buffer for b in buf_uops]
+    ubufs = [b.buffer for b in buf_uops if b.op is not Ops.BIND]
     metadata = si.arg.metadata
     if any(isinstance(x, MultiBuffer) for x in ubufs):
       assert all(isinstance(x, MultiBuffer) for x in ubufs), "kernel must all be multibuffer"

tinygrad/function.py

@@ -3,6 +3,7 @@ from typing import Generic, TypeVar, Callable, cast, overload
 from tinygrad.helpers import Context, dedup, getenv
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, PatternMatcher, UPat
 from tinygrad.tensor import Tensor
+from tinygrad.nn.state import get_state_dict
 
 def add_to_ctx(ctx, x:UOp):
   ret = x.param_like(len(ctx))
@@ -11,7 +12,7 @@ def add_to_ctx(ctx, x:UOp):
 
 pm_ctx = PatternMatcher([
   (UPat((Ops.BUFFER, Ops.BIND), name="x"), add_to_ctx),
-  (UPat((Ops.ASSIGN, Ops.CONTIGUOUS), name="x"),
+  (UPat((Ops.AFTER, Ops.CONTIGUOUS), name="x"),
    lambda ctx,x: add_to_ctx(ctx,x) if not x.op_in_backward_slice_with_self(Ops.PARAM) else None),
   # strip UNIQUE from unique consts — they don't need buffer identity inside function bodies
   (UPat(Ops.CONST, src=(UPat(Ops.UNIQUE), UPat(Ops.DEVICE)), name="x"), lambda ctx,x: x.replace(src=(x.src[1],))),
@@ -19,39 +20,50 @@ pm_ctx = PatternMatcher([
 
 ReturnType = TypeVar('ReturnType')
 class _function(Generic[ReturnType]):
-  def __init__(self, fxn:Callable[..., ReturnType], *, precompile:bool=False):
+  def __init__(self, fxn:Callable[..., ReturnType], *, precompile:bool=False, precompile_backward:bool=False,
+               allow_implicit:bool=True, grad_fxn:Callable|None=None):
     self.fxn = fxn
     self.precompile = precompile
+    self.precompile_backward = precompile_backward
+    self.allow_implicit = allow_implicit
+    self.grad_fxn = grad_fxn
 
   def __get__(self, obj, objtype=None): return functools.partial(self.__call__, obj) if obj is not None else self
 
   def __call__(self, *args, **kwargs) -> ReturnType:
-    input_uops: list[UOp] = [(t.uop if isinstance(t, Tensor) else t)
-                             for name,t in list(enumerate(args))+sorted(kwargs.items()) if isinstance(t, (Tensor, UOp))]
-
-    # use the base
-    #input_uops = [x.multibase for x in input_uops]
+    params = get_state_dict((args, kwargs), tensor_type=(Tensor, UOp)).values()
 
     # deduplicate input_uops, keeping the first occurrence index for each unique uop
-    call_uops: list[UOp] = dedup(input_uops)
+    call_uops: list[UOp] = dedup([(t.uop if isinstance(t, Tensor) else t) for t in params])
 
     # disable realize/schedule while this is running
     # run it and do surgery later
     with Context(ALLOW_DEVICE_USAGE=getenv("DEVICE_IN_FUNCTION_BUG", 0)):
       ret = self.fxn(*args, **kwargs)
-    assert isinstance(ret, Tensor), "only supports one tensor return for now"
+    if isinstance(ret, Tensor):
+      uret = ret.uop
+    elif isinstance(ret, tuple) and all(isinstance(x, Tensor) for x in ret):
+      uret = UOp.maketuple(*[x.uop for x in ret])
+    else:
+      raise RuntimeError(f"function return type {type(ret)} not supported")
 
     # replace the known inputs with params (using deduplicated slots)
     subs = {}
     for i,x in enumerate(call_uops): subs[x] = x.param_like(i)
-    uret = ret.uop.substitute(subs)
+    uret = uret.substitute(subs)
 
     # add contiguous to call_uops
     #call_uops = [x.contiguous() for x in call_uops]
 
     # the BUFFERs that are left are the implicit inputs
+    num_explicit = len(call_uops)
     uret = graph_rewrite(uret, pm_ctx, call_uops, bottom_up=True, name="get_implicit_inputs")
     name = getattr(self.fxn, '__qualname__', None) or type(self.fxn).__qualname__
+    if not self.allow_implicit:
+      implicit_buffers = [x for x in call_uops[num_explicit:] if x.op is Ops.BUFFER]
+      if implicit_buffers:
+        buf_strs = '\n  '.join(f"{i}: dtype={b.dtype}, size={b.size}, device={b.device}" for i,b in enumerate(implicit_buffers))
+        raise RuntimeError(f"function {name} has {len(implicit_buffers)} implicit buffer(s), but allow_implicit=False\n  {buf_strs}")
 
     # assign output
     #pbuffer = uret.param_like(len(call_uops))
@@ -60,14 +72,23 @@ class _function(Generic[ReturnType]):
     #call = assigned.call(*call_uops, buffer, name=name)
     #ret = buffer.after(call)
 
-    ret = uret.call(*call_uops, name=name, precompile=self.precompile)
-    return cast(ReturnType, Tensor(ret, device=ret.device))
+    fret = uret.call(*call_uops, grad_fxn=self.grad_fxn, name=name, precompile=self.precompile,
+                     precompile_backward=self.precompile_backward)
+    if isinstance(ret, tuple):
+      return cast(ReturnType, tuple(Tensor(fret.gettuple(i), device=fret.device) for i in range(len(ret))))
+    else:
+      return cast(ReturnType, Tensor(fret.gettuple(0), device=fret.device))
 
 # overload signatures support both @function and @function(precompile=True) syntax
 @overload
-def function(fxn:Callable[..., ReturnType], *, precompile:bool=False) -> _function[ReturnType]: ...
+def function(fxn:Callable[..., ReturnType], *, precompile:bool=False, precompile_backward:bool=False,
+             allow_implicit:bool=True, grad_fxn:Callable|None=None) -> _function[ReturnType]: ...
 @overload
-def function(fxn:None=None, *, precompile:bool=False) -> Callable[[Callable[..., ReturnType]], _function[ReturnType]]: ...
-def function(fxn=None, *, precompile:bool=False):
-  if fxn is None: return lambda f: _function(f, precompile=precompile)
-  return _function(fxn, precompile=precompile)
+def function(fxn:None=None, *, precompile:bool=False, precompile_backward:bool=False,
+             allow_implicit:bool=True, grad_fxn:Callable|None=None) -> Callable[[Callable[..., ReturnType]], _function[ReturnType]]: ...
+def function(fxn=None, *, precompile:bool=False, precompile_backward:bool=False, allow_implicit:bool=True, grad_fxn:Callable|None=None):
+  if fxn is None:
+    return lambda f: _function(f, precompile=precompile, precompile_backward=precompile_backward,
+                               allow_implicit=allow_implicit, grad_fxn=grad_fxn)
+  return _function(fxn, precompile=precompile, precompile_backward=precompile_backward,
+                   allow_implicit=allow_implicit, grad_fxn=grad_fxn)

tinygrad/gradient.py

@@ -13,21 +13,33 @@ def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):
     return ((mask/broadcast_to_input(count)) * broadcast_to_input(ctx),)
   if op == Ops.MUL: return (broadcast_to_input(ctx * ret) / ret.src[0],)
 
-def call_gradient(ctx:UOp, k:UOp) -> tuple[UOp|None, ...]:
-  if k.arg.grad_fxn is not None: return (None,) + k.arg.grad_fxn(ctx, k)
-  # auto-differentiate the function
+def _compact_params(body:UOp, all_args:tuple[UOp, ...]) -> tuple[UOp, tuple[UOp, ...]]:
+  """Remove unused PARAMs from body and return compacted (body, args)."""
+  used = sorted({p.arg: p for p in body.toposort() if p.op is Ops.PARAM}.items())
+  return body.substitute({p: p.replace(arg=j) for j,(_, p) in enumerate(used)}, walk=True), tuple(all_args[i] for i,_ in used)
+
+def call_gradient(ctx:UOp, k:UOp, needed:set[int]) -> tuple[UOp|None, ...]:
   fxn, args = k.src[0], k.src[1:]
+  if k.arg.grad_fxn is not None:
+    if ctx.op is Ops.TUPLE:
+      real = [g for g in ctx.src if g.op is not Ops.NOOP]
+      return (None,) + (k.arg.grad_fxn(*real, call=k) if len(real) > 1 else k.arg.grad_fxn(real[0], k))
+    return (None,) + k.arg.grad_fxn(ctx, k)
+  assert fxn.op is Ops.TUPLE, f"expected TUPLE body for gradient, got {fxn.op}"
   params = {x.arg:x for x in fxn.toposort(enter_calls=False) if x.op == Ops.PARAM}
-  grads = compute_gradient(fxn, ctx.param_like(len(args)), set(params.values()))
-  ret: list[UOp|None] = [None]
-  for i in range(len(args)):
-    if (p:=params.get(i, None)) is not None and p in grads:
-      # TODO: compact the args and remove unused ones
-      assert not grads[p].op_in_backward_slice_with_self(Ops.BUFFER), "BUG: BUFFER in backward slice of grad"
-      ret.append(grads[p].call(*args, ctx, name=(k.arg.name or "")+f"_backward_{i}"))
-    else:
-      ret.append(None)
-  return tuple(ret)
+  grad_args = ctx.src
+  root_grad = UOp(Ops.TUPLE, src=tuple(fxn.src[i].const_like(0) if g.op is Ops.NOOP else g.param_like(len(args)+i) for i,g in enumerate(grad_args)))
+  grads = compute_gradient(fxn, root_grad, set(params.values()))
+  # for precompiled calls, substitute forward outputs with params so intermediates aren't recomputed
+  fwd_subs = {src: src.param_like(len(args)+len(grad_args)+i) for i, src in enumerate(fxn.src)} if k.arg.precompile else {}
+  fwd_outs = tuple(k.gettuple(i) for i in range(len(fxn.src))) if k.arg.precompile else ()
+  # collect needed gradient bodies, compact unused params, create a single backward CALL
+  grad_bodies = [(i, grads[p]) for i in needed if (p:=params.get(i)) is not None and p in grads]
+  bwd_body = UOp.maketuple(*(gb for _, gb in grad_bodies)).substitute(fwd_subs, walk=True)
+  bwd_body, compact_args = _compact_params(bwd_body, (*args, *grad_args, *fwd_outs))
+  bwd_call = bwd_body.call(*compact_args, name=(k.arg.name or "")+"_backward", precompile=k.arg.precompile_backward)
+  gb_map = {i: idx for idx, (i, _) in enumerate(grad_bodies)}
+  return (None,) + tuple(bwd_call.gettuple(gb_map[i]) if i in gb_map else None for i in range(len(args)))
 
 # ctx is grad_output
 pm_gradient = PatternMatcher([
@@ -56,26 +68,40 @@ pm_gradient = PatternMatcher([
   (UPat(Ops.FLIP, name="ret"), lambda ctx, ret: (ctx.flip([i for i,x in enumerate(ret.marg) if x]),)),
   (UPat(Ops.COPY, name="ret"), lambda ctx, ret: (ctx.copy_to_device(ret.src[0].device), None)),
   (UPat(Ops.MULTI, name="ret"), lambda ctx, ret: ctx.shard(ret.device, ret.axis).src),
+  (UPat(Ops.TUPLE), lambda ctx: ctx.src),
   # NOTE: this is only correct when the KERNEL has a single output
   (UPat(Ops.AFTER), lambda ctx: (ctx, ctx)),
-  # gradient on CALL: use provided grad_fxn or auto-differentiate
-  (UPat(Ops.CALL, name="k"), call_gradient),
   # there's no gradient for bitcast
   (UPat(Ops.BITCAST), lambda: (None,)),
 ])
 
-def _deepwalk(root:UOp, targets:set[UOp]) -> list[UOp]:
+def _deepwalk(root:UOp, targets:set[UOp]) -> tuple[list[UOp], dict[UOp, bool]]:
   # compute the target path (top down)
   in_target_path: dict[UOp, bool] = {}
   for u in root.toposort(): in_target_path[u] = any(x in targets or in_target_path[x] for x in u.src)
-  # don't flow through DETACH/ASSIGN or anything not in target path
-  return list(root.toposort(lambda node: node.op not in {Ops.DETACH, Ops.ASSIGN} and in_target_path[node]))
+  # don't flow through DETACH or anything not in target path
+  return list(root.toposort(lambda node: node.op is not Ops.DETACH and in_target_path[node])), in_target_path
 
 def compute_gradient(root:UOp, root_grad:UOp, targets:set[UOp]) -> dict[UOp, UOp]:
-  grads = {root: root_grad}
-  for t0 in reversed(_deepwalk(root, targets)):
+  walk, in_target_path = _deepwalk(root, targets)
+  grads: dict[UOp, UOp] = {root: root_grad}
+  for t0 in reversed(walk):
     if t0 not in grads: continue
-    lgrads: tuple[UOp|None, ...]|None = cast(tuple[UOp|None, ...]|None, pm_gradient.rewrite(t0, ctx=grads[t0]))
+    # GETTUPLE: accumulate gradient into a TUPLE UOp on the CALL, process when we hit the CALL
+    if t0.op is Ops.GETTUPLE:
+      k = t0.src[0]  # the CALL
+      assert k.op is Ops.CALL and k.src[0].op is Ops.TUPLE
+      n_outputs = len(k.src[0].src)
+      prev = grads[k].src if k in grads else tuple(UOp(Ops.NOOP) for _ in range(n_outputs))
+      grads[k] = UOp.maketuple(*(prev[i] + grads[t0] if i == t0.arg and prev[i].op is not Ops.NOOP else
+                                 grads[t0] if i == t0.arg else prev[i] for i in range(n_outputs)))
+      continue
+    # CALL: pass needed param set so backward only computes required gradients
+    if t0.op is Ops.CALL:
+      needed = {i for i, arg in enumerate(t0.src[1:]) if arg in targets or in_target_path.get(arg, False)}
+      lgrads:tuple[UOp|None, ...]|None = call_gradient(grads[t0], t0, needed)
+    else:
+      lgrads = cast(tuple[UOp|None, ...]|None, pm_gradient.rewrite(t0, ctx=grads[t0]))
     if lgrads is None: raise RuntimeError(f"failed to compute gradient for {t0.op}\n\nin {str(t0)[0:1000]}...")
     assert len(lgrads) == len(t0.src), f"got {len(lgrads)} gradient, expected {len(t0.src)}"
     for k,v in zip(t0.src, lgrads):
@@ -87,8 +113,4 @@ def compute_gradient(root:UOp, root_grad:UOp, targets:set[UOp]) -> dict[UOp, UOp
         # we add the backward metadata to everything new in the graph
         for bw_uop in v.toposort(lambda x: x not in (t0, *t0.src, grads[t0])):
           all_metadata[bw_uop] = all_metadata.get(bw_uop, ())+backward_metadata
-  # end any ranges on grads with a reduce sum
-  for k,v in grads.items():
-    if len(v.ranges):
-      grads[k] = v.reduce(*v.ranges, arg=Ops.ADD)
   return grads

tinygrad/helpers.py

@@ -195,7 +195,8 @@ CPU_COUNT = ContextVar("CPU_COUNT", max(1, len(os.sched_getaffinity(0)) if hasat
 CPU_CC, CPU_LLVM, CPU_LVP = ContextVar("CPU_CC", ""), ContextVar("CPU_LLVM", 0), ContextVar("CPU_LVP", 0)
 NV_CC, NV_PTX, NV_NAK, NV_NVCC = ContextVar("NV_CC", ""), ContextVar("NV_PTX", 0), ContextVar("NV_NAK", 0), ContextVar("NV_NVCC", 0)
 CUDA_CC, CUDA_PTX, CUDA_NVCC = ContextVar("CUDA_CC", ""), ContextVar("CUDA_PTX", 0), ContextVar("CUDA_NVCC", 0)
-NULL_IR3, NULL_NAK, NULL_ALLOW_COPYOUT = ContextVar("NULL_IR3", 0), ContextVar("NULL_NAK", 0), ContextVar("NULL_ALLOW_COPYOUT", 0)
+NULL_QCOMCL, NULL_IR3, NULL_NAK = ContextVar("NULL_QCOMCL", 0), ContextVar("NULL_IR3", 0), ContextVar("NULL_NAK", 0)
+NULL_ALLOW_COPYOUT = ContextVar("NULL_ALLOW_COPYOUT", 0)
 AMD_CC, AMD_LLVM, AMD_HIPCC  = ContextVar("AMD_CC", ""), ContextVar("AMD_LLVM", 0), ContextVar("AMD_HIPCC", 0)
 QCOM_CC, QCOM_IR3 = ContextVar("QCOM_CC", ""), ContextVar("QCOM_IR3", 0)
 # VIZ implies PROFILE, but you can run PROFILE without VIZ
@@ -423,7 +424,9 @@ def fetch(url:str, name:pathlib.Path|str|None=None, subdir:str|None=None, gunzip
 
 def system(cmd:str, **kwargs) -> str:
   st = time.perf_counter()
-  ret = subprocess.check_output(cmd.split(), **kwargs).decode().strip()
+  try: ret = subprocess.check_output(cmd.split(), stderr=subprocess.STDOUT, **kwargs).decode().strip()
+  except subprocess.CalledProcessError as e:
+    raise RuntimeError(f"system: '{cmd}' failed with exit code {e.returncode}\n{(e.output or b'').decode().strip()}") from e
   if DEBUG >= 1: print(f"system: '{cmd}' returned {len(ret)} bytes in {(time.perf_counter() - st)*1e3:.2f} ms")
   return ret
 

tinygrad/nn/__init__.py

@@ -303,16 +303,22 @@ class RMSNorm:
     x = self._norm(x.float()).cast(x.dtype)
     return x if self.weight is None else x * self.weight
 
-from tinygrad.uop.ops import UOp, KernelInfo, Ops
+from tinygrad.uop.ops import UOp, KernelInfo, Ops, AxisType
 def _embedding_bwd(grad_emb:UOp, call:UOp) -> tuple:
   weight, idx = call.src[1:]
-  # for multi-device: unshard inputs to one device
+  is_vocab_sharded = isinstance(weight.device, tuple) and weight.axis == 0
+  # for multi-device: replicate grad_emb and idx on all devices
   if isinstance(weight.device, tuple):
-    assert weight.axis is None, "sharded weights on Embedding not supported with USE_ATOMICS"
+    assert weight.axis is None or weight.axis == 0, "only vocab (axis=0) sharding supported on Embedding with USE_ATOMICS"
     grad_emb = grad_emb.copy_to_device(weight.device)
     idx = idx.copy_to_device(weight.device)
-  # weight is replicated, grad_weight should match
-  grad_weight_uop = Tensor.empty(weight.shape, dtype=dtypes.float, device=weight.device).uop
+  if is_vocab_sharded:
+    ndev = len(weight.device)
+    local_vocab_size = weight.shape[0] // ndev
+    grad_weight_uop = Tensor.empty(local_vocab_size, weight.shape[1], dtype=dtypes.float, device=weight.device).uop.multi(axis=0)
+  else:
+    # weight is replicated (or single device), grad_weight should match
+    grad_weight_uop = Tensor.empty(weight.shape, dtype=dtypes.float, device=weight.device).uop
 
   # TODO: how do we remove this dumb kernel and use Tensor.zeros?
   def _zero_kernel(out:UOp) -> UOp:
@@ -326,15 +332,35 @@ def _embedding_bwd(grad_emb:UOp, call:UOp) -> tuple:
   # this is the real atomic kernel
   def _embedding_bwd_kernel(grad_weight:UOp, grad_emb:UOp, idx:UOp) -> UOp:
     idx_flat, grad_emb_flat = idx.flatten(), grad_emb.reshape((idx.size, grad_weight.shape[-1]))
-    i = UOp.range(grad_emb_flat.shape[0], 0)  # batch_size * sequence_length
-    j = UOp.range(grad_emb_flat.shape[1], 1)  # embed_size
-    token_id = idx_flat[i].clip(0, grad_weight.shape[0]-1).cast(dtypes.index)
+
+    embed_size = grad_weight.shape[-1]
+    BLOCK_J = min(256, embed_size)
+    assert embed_size % BLOCK_J == 0, f"embed_size {embed_size} must be divisible by {BLOCK_J}"
+
+    n_j_blocks = embed_size // BLOCK_J
+    i = UOp.range(grad_emb_flat.shape[0], 0)         # batch_size * sequence_length -> GLOBAL
+    j_inner = UOp.range(BLOCK_J, 2, AxisType.LOOP if device in ("CPU", "NULL") else AxisType.LOCAL)  # BLOCK_J threads per workgroup
+    j_outer = UOp.range(n_j_blocks, 1)
+    j = j_outer * BLOCK_J + j_inner
+
+    if is_vocab_sharded:
+      # each device owns [offset, offset+local_vocab_size) of the global vocabulary
+      dnum = UOp.variable("_device_num", 0, ndev-1)
+      offset = dnum * local_vocab_size
+      global_token_id = idx_flat[i].cast(dtypes.index)
+      local_token_id = (global_token_id - offset).clip(0, grad_weight.shape[0]-1)
+      in_range = (global_token_id >= offset) & (global_token_id < (offset + local_vocab_size))
+      grad_val = in_range.where(grad_emb_flat[i, j].cast(dtypes.float), 0.0)
+    else:
+      local_token_id = idx_flat[i].clip(0, grad_weight.shape[0]-1).cast(dtypes.index)
+      grad_val = grad_emb_flat[i, j].cast(dtypes.float)
     # atomic scatter-add: grad_weight[token_id, j] += grad_emb_flat[i, j]
     if device in ("CPU", "NULL"): atomic_arg = "__atomic_fetch_add({0}, {1}, __ATOMIC_RELAXED);"
     elif device == "AMD": atomic_arg = "__hip_atomic_fetch_add({0}, {1}, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);"
     else: raise NotImplementedError(f"no atomics for device {device}")
-    atomic = UOp(Ops.CUSTOM, dtypes.void, (grad_weight.index(token_id, j, ptr=True), grad_emb_flat[i, j].cast(dtypes.float)), arg = atomic_arg)
-    return atomic.end(i, j).sink(arg=KernelInfo(name="embedding_bwd", opts_to_apply=()))
+    atomic = UOp(Ops.CUSTOM, dtypes.void, (grad_weight.index(local_token_id, j, ptr=True), grad_val), arg = atomic_arg)
+    return atomic.end(i, j_outer, j_inner).sink(arg=KernelInfo(name="embedding_bwd", opts_to_apply=()))
+
   grad_weight_uop = grad_weight_uop.custom_kernel(grad_emb, idx, fxn=_embedding_bwd_kernel)[0]
 
   return (grad_weight_uop.cast(weight.dtype), None)

tinygrad/nn/onnx.py

@@ -153,21 +153,19 @@ class OnnxPBParser:
 
   def _parse_ModelProto(self) -> dict:
     """Entry point for parsing the ONNX model."""
-    obj: dict[str, Any] = {"opset_import": []}
+    graph: dict|None = None
+    opset_imports: list[OpSetId] = []
     for fid, wire_type in self._parse_message(self.reader.len):
       match fid:
-        case 4: obj["domain"] = self.reader.read_string()
-        case 5: obj["model_version"] = self.reader.read_int64()
-        case 7: obj["graph"] = self._parse_GraphProto()
-        case 8: obj["opset_import"].append(self._parse_OperatorSetIdProto())
+        case 7: graph = self._parse_GraphProto()
+        case 8: opset_imports.append(self._parse_OperatorSetIdProto())
         case _: self.reader.skip_field(wire_type)
-
+    assert graph is not None
     # update opset version
-    opset_imports = {Domain.from_onnx(x.get('domain')):x.get('version', 1) for x in obj["opset_import"]}
-    for n in obj["graph"]["node"]:
-      n_ = n["parsed_node"]
-      n["parsed_node"] = OnnxNode(n_.op, OpSetId(n_.opset_id.domain, opset_imports.get(n_.opset_id.domain, 1)), n_.inputs, n_.outputs, n_.opts)
-    return obj
+    versions = {opset.domain: opset.version for opset in opset_imports}
+    graph["node"] = [OnnxNode(n.op, OpSetId(n.opset_id.domain, versions.get(n.opset_id.domain, 1)), n.inputs, n.outputs, n.opts)
+                     for n in graph["node"]]
+    return graph
 
   def _parse_GraphProto(self) -> dict:
     obj: dict[str, Any] = {"node": [], "initializer": [], "input": [], "output": []}
@@ -181,26 +179,23 @@ class OnnxPBParser:
         case _: self.reader.skip_field(wire_type)
     return obj
 
-  def _parse_NodeProto(self) -> dict:
-    obj: dict[str, Any] = {"input": [], "output": [], "attribute": [], "domain": None}
+  def _parse_NodeProto(self) -> OnnxNode:
+    inputs: list[str] = []
+    outputs: list[str] = []
+    attributes: list[tuple[str, Any]] = []
+    domain: str|None = None
+    op_type = ""
     for fid, wire_type in self._parse_message(self._decode_end_pos()):
       match fid:
-        case 1: obj["input"].append(self.reader.read_string())
-        case 2: obj["output"].append(self.reader.read_string())
-        case 3: obj["name"] = self.reader.read_string()
-        case 4: obj["op_type"] = self.reader.read_string()
-        case 5: obj["attribute"].append(self._parse_AttributeProto())
-        case 6: obj["doc_string"] = self.reader.read_string()
-        case 7: obj["domain"] = self.reader.read_string()
+        case 1: inputs.append(self.reader.read_string())
+        case 2: outputs.append(self.reader.read_string())
+        case 4: op_type = self.reader.read_string()
+        case 5: attributes.append(self._parse_AttributeProto())
+        case 7: domain = self.reader.read_string()
         case _: self.reader.skip_field(wire_type)
+    return OnnxNode(op_type, OpSetId(Domain.from_onnx(domain), 1), tuple(inputs), tuple(outputs), dict(attributes))
 
-    # parse node
-    attributes = {attr_dict["name"]: attr_dict[AttributeType(attr_dict["type"]).to_field_name()] for attr_dict in obj["attribute"]}
-    opset_id = OpSetId(Domain.from_onnx(obj.get('domain')), 1)  # default version, to be updated later in _parse_ModelProto
-    obj["parsed_node"] = OnnxNode(obj["op_type"], opset_id, tuple(obj["input"]), tuple(obj["output"]), attributes)
-    return obj
-
-  def _parse_TensorProto(self) -> dict:
+  def _parse_TensorProto(self) -> tuple[str, Tensor]:
     obj: dict[str, Any] = {"dims": []}
     for fid, wire_type in self._parse_message(self._decode_end_pos()):
       match fid:
@@ -220,18 +215,16 @@ class OnnxPBParser:
     # load external data
     if self.load_external_data and obj.get("data_location", 0) == 1:
       if "external_data" not in obj: raise ValueError("no external_data")
-      location, length, offset = None, None, 0
-      for kv in obj["external_data"]:
-        if kv["key"] == "location": location = kv["value"]
-        elif kv["key"] == "offset": offset = int(kv["value"])
-        elif kv["key"] == "length": length = int(kv["value"])
-      if location is None: raise ValueError("no location in external_data")
+      ext = dict(obj["external_data"])
+      if "location" not in ext: raise ValueError("no location in external_data")
+      offset = int(ext.get("offset", "0"))
+      length = int(ext["length"]) if "length" in ext else None
 
       if self.file_path is None:
         if isinstance(self.tensor.device, str) and self.tensor.device.startswith("DISK:"):
           self.file_path = pathlib.Path(self.tensor.device[5:])
         else: raise ValueError("onnx external_data needs the origin file path, try passing onnx file path to onnx_load")
-      ext_path = self.file_path.parent.joinpath(location)
+      ext_path = self.file_path.parent.joinpath(ext["location"])
       if not ext_path.exists(): raise FileNotFoundError(f"external location not exists: {ext_path}")
 
       ext_tensor = Tensor(ext_path)
@@ -241,23 +234,20 @@ class OnnxPBParser:
     # parse tensor
     to_dtype = dtype_fallback(true_dtype := OnnxDataType(obj['data_type']).to_dtype(), "buffer parse")
     shape = tuple(obj['dims'])
-    present_fields = [field for field in ['float_data', 'int32_data', 'int64_data', 'double_data', 'uint64_data', 'raw_data'] if field in obj]
-    assert len(present_fields) == 1, f"only 1 data field is allowed from {obj=}"
-    data = obj[present_fields[0]]
-    if not isinstance(data, Tensor):
-      obj["parsed_tensor"] = Tensor(data, dtype=to_dtype).reshape(shape)
-      return obj
-    assert isinstance(data, Tensor) and data.dtype == dtypes.uint8, data
+    data_fields = [f for f in ('float_data','int32_data','int64_data','double_data','uint64_data','raw_data') if f in obj]
+    data = obj[get_single_element(data_fields)]
+    name = obj.get("name", "")
+    if not isinstance(data, Tensor): return name, Tensor(data, dtype=to_dtype).reshape(shape)
+    assert data.dtype == dtypes.uint8, data
     data = data.bitcast(true_dtype).reshape(shape)
     data = data.to(Device.DEFAULT) if true_dtype is to_dtype else data.to("cpu").cast(to_dtype).to(Device.DEFAULT)
     # const folding
     if shape == ():
       if data.dtype == dtypes.float16 and sys.version_info < (3, 12): data = data.cast(dtypes.float32)
       data = Tensor(data.item(), dtype=to_dtype).reshape(shape)
-    obj["parsed_tensor"] = data
-    return obj
+    return name, data
 
-  def _parse_AttributeProto(self) -> dict:
+  def _parse_AttributeProto(self) -> tuple[str, Any]:
     obj: dict[str, Any] = {"floats": [], "ints": [], "strings": []}
     for fid, wire_type in self._parse_message(self._decode_end_pos()):
       match fid:
@@ -265,7 +255,7 @@ class OnnxPBParser:
         case 2: obj["f"] = self.reader.read_float()
         case 3: obj["i"] = self.reader.read_int64()
         case 4: obj["s"] = self.reader.read_bytes().data().tobytes().decode("utf8")
-        case 5: obj["t"] = self._parse_TensorProto()['parsed_tensor']
+        case 5: obj["t"] = self._parse_TensorProto()[1]
         case 6: obj["g"] = OnnxRunner._from_subgraph(self._parse_GraphProto())
         case 7: obj["floats"].append(self.reader.read_float())
         case 8: obj["ints"].append(self.reader.read_int64())
@@ -273,26 +263,22 @@ class OnnxPBParser:
         case 20: obj["type"] = self.reader.read_int64()
         case _: self.reader.skip_field(wire_type)
     obj["floats"], obj["ints"], obj["strings"] = tuple(obj["floats"]), tuple(obj["ints"]), tuple(obj["strings"])
-    return obj
+    return obj["name"], obj[AttributeType(obj["type"]).to_field_name()]
 
-  def _parse_ValueInfoProto(self) -> dict:
-    obj: dict[str, Any] = {}
+  def _parse_ValueInfoProto(self) -> tuple[str, OnnxValue|None]:
+    name, type_obj = "", None
     for fid, wire_type in self._parse_message(self._decode_end_pos()):
       match fid:
-        case 1: obj["name"] = self.reader.read_string()
-        case 2: obj["type"] = self._parse_TypeProto()
+        case 1: name = self.reader.read_string()
+        case 2: type_obj = self._parse_TypeProto()
         case _: self.reader.skip_field(wire_type)
-
-    # parse type
-    if "type" not in obj: return {**obj, "parsed_type": None}
-    type_obj = obj["type"]
+    if type_obj is None: return name, None
     if is_optional := "optional_type" in type_obj: type_obj = type_obj["optional_type"]["elem_type"]
     if is_sequence := "sequence_type" in type_obj: type_obj = type_obj["sequence_type"]["elem_type"]
     assert "tensor_type" in type_obj, type_obj
     shape_dims = type_obj['tensor_type'].get('shape', {}).get('dim', [])
-    obj['parsed_type'] = OnnxValue(tuple(d.get('dim_param') or d.get('dim_value') for d in shape_dims),
-                                   OnnxDataType(type_obj['tensor_type']['elem_type']).to_dtype(), is_optional, is_sequence)
-    return obj
+    return name, OnnxValue(tuple(d.get('dim_param') or d.get('dim_value') for d in shape_dims),
+                           OnnxDataType(type_obj['tensor_type']['elem_type']).to_dtype(), is_optional, is_sequence)
 
   def _parse_TypeProto(self) -> dict:
     obj: dict[str, Any] = {}
@@ -338,23 +324,24 @@ class OnnxPBParser:
         case _: self.reader.skip_field(wire_type)
     return obj
 
-  def _parse_StringStringEntryProto(self) -> dict:
-    obj: dict[str, Any] = {}
+  def _parse_StringStringEntryProto(self) -> tuple[str, str]:
+    key, value = "", ""
     for fid, wire_type in self._parse_message(self._decode_end_pos()):
       match fid:
-        case 1: obj["key"] = self.reader.read_string()
-        case 2: obj["value"] = self.reader.read_string()
+        case 1: key = self.reader.read_string()
+        case 2: value = self.reader.read_string()
         case _: self.reader.skip_field(wire_type)
-    return obj
+    return key, value
 
-  def _parse_OperatorSetIdProto(self) -> dict:
-    obj: dict[str, Any] = {}
+  def _parse_OperatorSetIdProto(self) -> OpSetId:
+    domain: str|None = None
+    version = 1
     for fid, wire_type in self._parse_message(self._decode_end_pos()):
       match fid:
-        case 1: obj["domain"] = self.reader.read_string()
-        case 2: obj["version"] = self.reader.read_int64()
+        case 1: domain = self.reader.read_string()
+        case 2: version = self.reader.read_int64()
         case _: self.reader.skip_field(wire_type)
-    return obj
+    return OpSetId(Domain.from_onnx(domain), version)
 
 # ***** python const *****
 required_input_python_consts: dict[str, tuple[int, ...]] = {
@@ -380,16 +367,15 @@ class OnnxRunner:
     model_path: The ONNX model, provided as a file path (a string or Path object) or a Tensor.
   """
   def __init__(self, model_path: Tensor | str | pathlib.Path):
-    model = OnnxPBParser(model_path, load_external_data=True).parse()
-    self._init_from_graph(model["graph"])
+    self._init_from_graph(OnnxPBParser(model_path, load_external_data=True).parse())
 
   def _init_from_graph(self, graph: dict, is_subgraph: bool = False):
-    self.is_training = any(n['parsed_node'].opset_id.domain in {Domain.AI_ONNX_TRAINING, Domain.AI_ONNX_PREVIEW_TRAINING} for n in graph["node"])
+    self.is_training = any(n.opset_id.domain in {Domain.AI_ONNX_TRAINING, Domain.AI_ONNX_PREVIEW_TRAINING} for n in graph["node"])
     self.graph_name = graph["name"] if is_subgraph else ""
-    self.graph_values = {"": None, **{i["name"]: i["parsed_tensor"] for i in graph["initializer"]}}
-    self.graph_inputs = {i["name"]: i["parsed_type"] for i in graph["input"] if i["name"] not in self.graph_values}
-    self.graph_outputs = tuple(o["name"] for o in graph["output"])
-    self.graph_nodes = tuple(n["parsed_node"] for n in graph["node"])
+    self.graph_values: dict[str, Any] = {"": None, **dict(graph["initializer"])}
+    self.graph_inputs = {name: typ for name, typ in graph["input"] if name not in self.graph_values}
+    self.graph_outputs = tuple(name for name, _ in graph["output"])
+    self.graph_nodes = tuple(graph["node"])
     # track names from initializers and Constant nodes for fast path optimizations
     self.const_names: set[str] = set(self.graph_values.keys()) | {o for n in self.graph_nodes if n.op == "Constant" for o in n.outputs}
 
@@ -511,7 +497,7 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
     o_ = [((i - 1) // s + 1) for i,s in zip(i_, s_)]
     return _onnx_pads_to_tiny_pads(_auto_pad([(o-1)*s+k-i for o,i,k,s in zip(o_, i_, k_, s_)], auto_pad))
 
-  def _clamp_cast(x:Tensor, dtype:DType): return x.clamp(dtypes.min(dtype), dtypes.max(dtype)).cast(dtype)
+  def _clamp_cast(x:Tensor, dtype:DType): return x.clamp(dtype.min, dtype.max).cast(dtype)
 
   def _prepare_quantize(x:Tensor, scale:Tensor, zero_point:Tensor|int, axis=1, block_size=0):
     if axis < 0: axis += x.ndim
@@ -1223,7 +1209,7 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
 
   def DynamicQuantizeLinear(x: Tensor):
     # only support uint8
-    qmin, qmax = dtypes.min(dtypes.uint8), dtypes.max(dtypes.uint8)
+    qmin, qmax = dtypes.uint8.min, dtypes.uint8.max
     scale = (x.max().maximum(0) + ((-x).max()).maximum(0)) / (qmax - qmin)
     zero_point = _clamp_cast((qmin - x.min() / scale).round(), dtypes.uint8)
     y = _clamp_cast((x / scale).round() + zero_point, dtypes.uint8)

tinygrad/nn/state.py

@@ -336,15 +336,14 @@ def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
       d = blocks[:,-2:].bitcast(dtypes.float16).cast(dtypes.float32).expand((-1, 256))
       return d * (xl.bitwise_or(xh).bitcast(dtypes.int8) - 32).flatten(-2) * scales
     if ggml_type == 39:
-      e_int = blocks[:, 0].cast(dtypes.int32)
-      d = ((e_int >= 2).cast(dtypes.float32) * (e_int.cast(dtypes.float32) - 128).exp2() +
-           (e_int == 1).cast(dtypes.float32) * 2.0**(-127) +
-           (e_int == 0).cast(dtypes.float32) * 2.0**(-128)).unsqueeze(-1)
+      e = blocks[:, 0].cast(dtypes.uint32)
+      small_bits = Tensor([0x00200000, 0x00400000], dtype=dtypes.uint32, device=t.device)[e.clip(0, 1).cast(dtypes.int32)] # e = 0 or e = 1 case
+      d = (e < 2).where(small_bits, ((e - 1) * 0x00800000).cast(dtypes.uint32)).bitcast(dtypes.float32).unsqueeze(-1)
       codes = q_to_uint8(blocks[:, 1:17], 4)
-      sign = 1.0 - codes.rshift(3).cast(dtypes.float32) * 2.0
-      exp, mant = codes.rshift(1).bitwise_and(0x3).cast(dtypes.float32), codes.bitwise_and(0x1).cast(dtypes.float32)
-      fp4_val = sign * 2.0 * ((exp != 0).cast(dtypes.float32) * (1.0 + 0.5 * mant) * (exp - 1.0).exp2() +
-                              (exp == 0).cast(dtypes.float32) * 0.5 * mant)
+      fp4_lut = Tensor([0.0, 1.0, 2.0, 3.0, 4.0, 6.0, 8.0, 12.0,
+                       -0.0,-1.0,-2.0,-3.0,-4.0,-6.0,-8.0,-12.0],
+                       dtype=dtypes.float32, device=t.device)
+      fp4_val = fp4_lut[codes]
       return (fp4_val * d).flatten(-2)[:n]
   raise ValueError(f"GGML type '{ggml_type}' is not supported!")