Merge branch 'master' into add_local_buffer

2025-10-06 08:58:46 +08:00
parent b9f7a7e218 1216fff781
commit df1b379a36
31 changed files with 538 additions and 128 deletions
@@ -109,7 +109,7 @@ jobs:
    - name: Train MNIST
      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=320 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=330 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
      run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=385 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
    #- name: Run 10 CIFAR training steps w BF16
@@ -452,6 +452,8 @@ jobs:
        run: CL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
      - name: Test MLPerf stuff
        run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
+      - name: NULL=1 beautiful_mnist_multigpu
+        run: MAX_BUFFER_SIZE=0 NULL=1 python examples/beautiful_mnist_multigpu.py
      - name: Test Bert training
        run: MAX_BUFFER_SIZE=0 NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
@@ -511,6 +511,33 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
      # happens with BENCHMARK set
      pass

+# stable diffusion callbacks to match mlperf ref; declared here because they're pickled
+def filter_dataset(sample:dict): return {k:v for k,v in sample.items() if k in {'npy', 'txt'}}
+def collate(batch:list[dict]):
+  ret = {"npy": [], "txt": [], "__key__": []}
+  for sample in batch:
+    for k,v in sample.items():
+      ret[k].append(v)
+  return ret
+def collate_fn(batch): return batch
+
+# Reference (code): https://github.com/mlcommons/training/blob/2f4a93fb4888180755a8ef55f4b977ef8f60a89e/stable_diffusion/ldm/data/webdatasets.py, Line 55
+# Reference (params): https://github.com/mlcommons/training/blob/ab4ae1ca718d7fe62c369710a316dff18768d04b/stable_diffusion/configs/train_01x08x08.yaml, Line 107
+def batch_load_train_stable_diffusion(urls:str, BS:int):
+  import webdataset
+  dataset = webdataset.WebDataset(urls=urls, resampled=True, cache_size=-1, cache_dir=None)
+  dataset = dataset.shuffle(size=1000)
+  dataset = dataset.decode()
+  dataset = dataset.map(filter_dataset)
+  dataset = dataset.batched(BS, partial=False, collation_fn=collate)
+  dataset = webdataset.WebLoader(dataset, batch_size=None, shuffle=False, num_workers=1, persistent_workers=True, collate_fn=collate_fn)
+
+  for x in dataset:
+    assert isinstance(x, dict) and all(isinstance(k, str) for k in x.keys()) and all(isinstance(v, list) for v in x.values())
+    assert all(isinstance(moment_mean_logvar, np.ndarray) and moment_mean_logvar.shape==(1,8,64,64) for moment_mean_logvar in x["npy"])
+    assert all(isinstance(caption, str) for caption in x["txt"])
+    yield x
+
 # llama3

 class BinIdxDataset:
@@ -1493,6 +1493,144 @@ def train_llama3():
          safe_save(get_state_dict(model), fn)
        break

+def train_stable_diffusion():
+  from extra.models.unet import UNetModel
+  from examples.mlperf.dataloader import batch_load_train_stable_diffusion
+  from examples.mlperf.lr_schedulers import LambdaLR, LambdaLinearScheduler
+  from examples.mlperf.initializers import init_stable_diffusion
+  from examples.mlperf.helpers import get_training_state
+  import numpy as np
+
+  config = {}
+  GPUS               = config["GPUS"]                   = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 1))]
+  seed               = config["seed"]                   = getenv("SEED", 12345)
+  # ** hyperparameters **
+  BS                 = config["BS"]                     = getenv("BS", 1 * len(GPUS))
+  BASE_LR            = config["LEARNING_RATE"]          = getenv("LEARNING_RATE", 2.5e-7)
+  # https://github.com/mlcommons/training_policies/blob/cfa99da479b8d5931f7a3c67612d021dfb47510a/training_rules.adoc#benchmark_specific_rules
+  # "Checkpoint must be collected every 512,000 images. CEIL(512000 / global_batch_size) if 512000 is not divisible by GBS."
+  # NOTE: It's inferred that "steps" is the unit for the output of the CEIL formula, based on all other cases of CEIL in the rules
+  CKPT_STEP_INTERVAL = config["CKPT_STEP_INTERVAL"]     = getenv("CKPT_STEP_INTERVAL", math.ceil(512_000 / BS))
+  CKPTDIR            = config["CKPTDIR"]                = Path(getenv("CKPTDIR", "./checkpoints"))
+  DATADIR            = config["DATADIR"]                = Path(getenv("DATADIR", "./datasets"))
+  UNET_CKPTDIR       = config["UNET_CKPTDIR"]           = Path(getenv("UNET_CKPTDIR", "./checkpoints"))
+  TOTAL_CKPTS        = config["TOTAL_CKPTS"]            = getenv("TOTAL_CKPTS", 0)
+
+  print(f"training on {GPUS}")
+  lr = BS * BASE_LR
+  print(f"BS={BS}, BASE_LR={BASE_LR}, lr={lr}")
+  print(f"CKPT_STEP_INTERVAL = {CKPT_STEP_INTERVAL}")
+  for x in GPUS: Device[x]
+  if (WANDB := getenv("WANDB", "")):
+    import wandb
+    wandb.init(config=config, project="MLPerf-Stable-Diffusion")
+
+  Tensor.manual_seed(seed)  # seed for weight initialization
+  model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod = init_stable_diffusion("v2-mlperf-train", CKPTDIR / "sd" / "512-base-ema.ckpt", GPUS)
+
+  optimizer = AdamW(get_parameters(unet))
+  lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
+  lr_scheduler = LambdaLR(optimizer, Tensor(lr, dtype=dtypes.float, device=optimizer.device), lambda_lr_callback)
+
+  @TinyJit
+  def train_step(mean:Tensor, logvar:Tensor, tokens:Tensor, unet:UNetModel, optimizer:LAMB, lr_scheduler:LambdaLR) -> Tensor:
+    optimizer.zero_grad()
+
+    timestep = Tensor.randint(BS, low=0, high=model.alphas_cumprod.shape[0], dtype=dtypes.int, device=GPUS[0])
+    latent_randn = Tensor.randn(*mean.shape, device=GPUS[0])
+    noise = Tensor.randn(*mean.shape, device=GPUS[0])
+    for t in (mean, logvar, tokens, timestep, latent_randn, noise):
+      t.shard_(GPUS, axis=0)
+
+    std = Tensor.exp(0.5 * logvar.clamp(-30.0, 20.0))
+    latent = (mean + std * latent_randn) * 0.18215
+
+    sqrt_alphas_cumprod_t = sqrt_alphas_cumprod[timestep].reshape(timestep.shape[0], 1, 1, 1)
+    sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alphas_cumprod[timestep].reshape(timestep.shape[0], 1, 1, 1)
+    latent_with_noise = sqrt_alphas_cumprod_t * latent + sqrt_one_minus_alphas_cumprod_t * noise
+    v_true = sqrt_alphas_cumprod_t * noise - sqrt_one_minus_alphas_cumprod_t * latent
+
+    context = model.cond_stage_model.embed_tokens(tokens)
+
+    out = unet(latent_with_noise, timestep, context)
+    loss = ((out - v_true) ** 2).mean()
+    del mean, logvar, std, latent, noise, sqrt_alphas_cumprod_t, sqrt_one_minus_alphas_cumprod_t
+    del out, v_true, context, latent_randn, tokens, timestep
+    loss.backward()
+
+    optimizer.step()
+    lr_scheduler.step()
+    loss, out_lr = loss.detach().to("CPU"), optimizer.lr.to("CPU")
+    Tensor.realize(loss, out_lr)
+    return loss, out_lr
+    
+  # checkpointing takes ~9 minutes without this, and ~1 minute with this
+  @TinyJit
+  def ckpt_to_cpu():
+    ckpt = get_training_state(unet, optimizer, lr_scheduler)
+    # move to CPU first so more GPU bufs aren't created (can trigger OOM)
+    for k,v in ckpt.items(): ckpt[k] = v.detach().to("CPU")
+    Tensor.realize(*[v for v in ckpt.values()])
+    for k,v in ckpt.items(): ckpt[k] = v.cast(v.dtype.base).contiguous()
+    Tensor.realize(*[v for v in ckpt.values()])
+    return ckpt
+
+  # training loop
+  dl = batch_load_train_stable_diffusion(f'{DATADIR}/laion-400m/webdataset-moments-filtered/{{00000..00831}}.tar', BS)
+  # for tests
+  saved_checkpoints = []
+
+  train_start_time = time.perf_counter()
+  t0 = t6 = time.perf_counter()
+  for i, batch in enumerate(dl, start=1):
+    loop_time = time.perf_counter() - t0
+    t0 = time.perf_counter()
+    dl_time = t0 - t6
+    GlobalCounters.reset()
+
+    mean, logvar = np.split(np.concatenate(batch["npy"], axis=0), 2, axis=1)
+    mean, logvar = Tensor(mean, dtype=dtypes.float32, device="CPU"), Tensor(logvar, dtype=dtypes.float32, device="CPU")
+    tokens = []
+    for text in batch['txt']: tokens += model.cond_stage_model.tokenizer.encode(text, pad_with_zeros=True)
+    tokens = Tensor(tokens, dtype=dtypes.int32, device="CPU").reshape(-1, 77)
+
+    t1 = time.perf_counter()
+    loss, lr = train_step(mean, logvar, tokens, unet, optimizer, lr_scheduler)
+    loss_item, lr_item = loss.item(), lr.item()
+    t2 = time.perf_counter()
+
+    if i == 3:
+      for _ in range(3): ckpt_to_cpu() # do this at the beginning of run to prevent OOM surprises when checkpointing
+      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
+      
+    total_train_time = time.perf_counter() - train_start_time
+    if WANDB:
+      wandb.log({"train/loss": loss_item, "train/lr": lr_item, "train/loop_time_prev": loop_time, "train/dl_time": dl_time, "train/step": i,
+                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (t2-t1), "train/input_prep_time": t1-t0,
+                    "train/train_step_time": t2-t1, "train/total_time": total_train_time})
+
+      if i == 1 and wandb.run is not None:
+        with open(f"{UNET_CKPTDIR}/wandb_run_id_{wandb.run.id}", "w") as f:
+          f.write(f"wandb.run.id = {wandb.run.id}")
+
+    if i % CKPT_STEP_INTERVAL == 0:
+      # https://github.com/mlcommons/training_policies/blob/cfa99da479b8d5931f7a3c67612d021dfb47510a/training_rules.adoc#benchmark_specific_rules
+      # "evaluation is done offline, the time is not counted towards the submission time."
+      fn = f"{UNET_CKPTDIR}/{i}.safetensors"
+      print(f"saving unet checkpoint at {fn}")
+      saved_checkpoints.append(fn)
+      safe_save({k.replace("model.", ""):v for k,v in ckpt_to_cpu().items() if k.startswith("model.")}, fn)
+      if TOTAL_CKPTS and i == TOTAL_CKPTS * CKPT_STEP_INTERVAL:
+        print(f"ending run after {i} steps ({TOTAL_CKPTS} checkpoints collected)")
+        return saved_checkpoints
+
+    t3 = time.perf_counter()
+    print(f"""step {i}: {GlobalCounters.global_ops * 1e-9 / (t2-t1):9.2f} GFLOPS, mem_used: {GlobalCounters.mem_used / 1e9:.2f} GB,
+  loop_time_prev: {loop_time:.2f}, dl_time: {dl_time:.2f}, input_prep_time: {t1-t0:.2f}, train_step_time: {t2-t1:.2f},
+  t3-t2: {t3-t2:.4f}, loss:{loss_item:.5f}, lr:{lr_item:.3e}, total_train_time:{total_train_time:.2f}
+  """)
+    t6 = time.perf_counter()
+
 if __name__ == "__main__":
  multiprocessing.set_start_method('spawn')

@@ -1501,7 +1639,7 @@ if __name__ == "__main__":
  else: bench_log_manager = contextlib.nullcontext()

  with Tensor.train():
-    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn").split(","):
+    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
      nm = f"train_{m}"
      if nm in globals():
        print(f"training {m}")
@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+
+DATETIME=${2:-$(date "+%m%d%H%M")}
+LOGFILE="${HOME}/logs/sd_mi300x_${DATETIME}.log"
+# UNET_CKPTDIR must be set: training saves checkpoints to this path, then a separate eval process scans this path to know which checkpoints to eval
+export UNET_CKPTDIR="${HOME}/stable_diffusion/training_checkpoints/${DATETIME}"
+mkdir -p "${HOME}/logs" "$UNET_CKPTDIR"
+
+# run this script in isolation when using the --bg flag
+if [[ "${1:-}" == "--bg" ]]; then
+  echo "logging output to $LOGFILE"
+  echo "saving UNet checkpoints to $UNET_CKPTDIR"
+  script_path="$(readlink -f "${BASH_SOURCE[0]}")"
+  nohup bash "$script_path" run "$DATETIME" >"$LOGFILE" 2>&1 & disown $!
+  exit 0
+fi
+
+# venv management
+if [[ -d .venv-sd-mlperf ]]; then
+  . .venv-sd-mlperf/bin/activate
+else
+  python3 -m venv .venv-sd-mlperf && . .venv-sd-mlperf/bin/activate
+  pip install --index-url https://download.pytorch.org/whl/cpu torch && pip install tqdm numpy ftfy regex pillow scipy wandb webdataset
+fi
+pip list
+apt list --installed | grep amdgpu
+rocm-smi --version
+modinfo amdgpu | grep version
+
+export BEAM=2 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 IGNORE_JIT_FIRST_BEAM=1 HCQDEV_WAIT_TIMEOUT_MS=300000
+export AMD_LLVM=0 # bf16 seems to require this
+export DATADIR="/raid/datasets/stable_diffusion"
+export CKPTDIR="/raid/weights/stable_diffusion"
+export EVAL_CKPT_DIR=$UNET_CKPTDIR
+export MODEL="stable_diffusion" PYTHONPATH="."
+export GPUS=8 BS=304
+export CONTEXT_BS=816 DENOISE_BS=600 DECODE_BS=384 INCEPTION_BS=560 CLIP_BS=240
+export WANDB=1
+export PARALLEL=4
+export PYTHONUNBUFFERED=1
+sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setperfdeterminism 1500 || exit 1
+
+# Retry BEAM search if script fails before BEAM COMPLETE is printed, but don't retry after that
+run_retry(){ local try=0 max=5 code tmp py pgid kids
+  while :; do
+    tmp=$(mktemp)
+    setsid bash -c 'exec env "$@"' _ "$@" > >(tee -a "$LOGFILE" | tee "$tmp") 2>&1 &
+    py=$!; pgid=$(ps -o pgid= -p "$py" | tr -d ' ')
+    wait "$py"; code=$?
+    [[ -n "$pgid" ]] && { kill -TERM -"$pgid" 2>/dev/null; sleep 1; kill -KILL -"$pgid" 2>/dev/null; }
+    kids=$(pgrep -P "$py" || true)
+    while [[ -n "$kids" ]]; do
+      kill -TERM $kids 2>/dev/null; sleep 0.5
+      kids=$(for k in $kids; do pgrep -P "$k" || true; done)
+    done
+    grep -q 'BEAM COMPLETE' "$tmp" && { rm -f "$tmp"; return 1; }
+    rm -f "$tmp"
+    ((code==0)) && return 0
+    ((try>=max)) && return 2
+    ((try++)); sleep 90; echo "try = ${try}"
+  done
+}
+
+# Power limiting to 400W is only needed if GPUs fall out of sync (causing 2.2x increased train time) at higher power, which has been observed at 450W
+sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setpoweroverdrive 750 && \
+run_retry TOTAL_CKPTS=7 python3 examples/mlperf/model_train.py; (( $? == 2 )) && { echo "training failed before BEAM completion"; exit 2; }
+sleep 90
+
+run_retry EVAL_SAMPLES=600 python3 examples/mlperf/model_eval.py; (( $? == 2 )) && { echo "eval failed before BEAM completion"; exit 2; }
+# Checkpoints will be evaluated in reverse chronological order, even if above training crashed early
+# STOP_IF_CONVERGED=1: Stop the eval after the first time convergence is detected; no more checkpoints will be evaluated after that.
+STOP_IF_CONVERGED=1 python3 examples/mlperf/model_eval.py
@@ -1,6 +1,6 @@
 [pytest]
 norecursedirs = extra
-timeout = 180
+timeout = 240
 timeout_method = thread
 timeout_func_only = true
 testpaths = test
@@ -221,7 +221,7 @@ class TestOpt(unittest.TestCase):
      for axis in [0, 1]:
        for n in [4, 8, 16]:
          b = torch.ones(n, n).sum(axis).reshape(n, 1).expand(n, n).sum(axis)
-          with CLCache(allowed=2):
+          with CLCache(allowed=3 if RANGEIFY else 2):
            a = Tensor.ones(n, n).contiguous().sum(axis).reshape(n, 1).expand(n, n).sum(axis)
            a.realize()
          np.testing.assert_allclose(a.numpy(), b.numpy(), rtol=1e-3, atol=1e-5)
@@ -231,7 +231,7 @@ class TestOpt(unittest.TestCase):
      axis1, axis2 = 0, 1
      for n in [4, 8, 16]:
        b = torch.ones(n, n).sum(axis1).reshape(n, 1).expand(n, n).sum(axis2)
-        with CLCache(allowed=2):
+        with CLCache(allowed=3 if RANGEIFY else 2):
          a = Tensor.ones(n, n).contiguous().sum(axis1).reshape(n, 1).expand(n, n).sum(axis2)
          a.realize()
        np.testing.assert_allclose(a.numpy(), b.numpy(), rtol=1e-3, atol=1e-5)
@@ -0,0 +1,23 @@
+import unittest, os
+from tempfile import TemporaryDirectory
+from tinygrad import Tensor
+from tinygrad.helpers import getenv
+from examples.mlperf.model_train import train_stable_diffusion
+
+class TestTrain(unittest.TestCase):
+  def test_train_to_ckpt(self):
+    # train for num_steps, save checkpoint, and stop training
+    num_steps = 42
+    os.environ.update({"MODEL": "stable_diffusion", "TOTAL_CKPTS": "1", "CKPT_STEP_INTERVAL": str(num_steps), "GPUS": "8", "BS": "304"})
+    # NOTE: update these based on where data/checkpoints are on your system
+    if not getenv("DATADIR", ""): os.environ["DATADIR"] = "/raid/datasets/stable_diffusion"
+    if not getenv("CKPTDIR", ""): os.environ["CKPTDIR"] = "/raid/weights/stable_diffusion"
+    with TemporaryDirectory(prefix="test-train") as tmp:
+      os.environ["UNET_CKPTDIR"] = tmp
+      with Tensor.train():
+        saved_ckpts = train_stable_diffusion()
+      expected_ckpt = f"{tmp}/{num_steps}.safetensors"
+      assert len(saved_ckpts) == 1 and saved_ckpts[0] == expected_ckpt
+
+if __name__=="__main__":
+  unittest.main()
@@ -176,7 +176,7 @@ class TestRealWorld(unittest.TestCase):
      for v in data.values(): v.to_(Device.DEFAULT)

      helper_test("train_bert", lambda: (data["input_ids"], data["segment_ids"], data["input_mask"], data["masked_lm_positions"], \
-          data["masked_lm_ids"], data["masked_lm_weights"], data["next_sentence_labels"]), train, 0.25, 357)
+          data["masked_lm_ids"], data["masked_lm_weights"], data["next_sentence_labels"]), train, 0.28, 357)

 if __name__ == '__main__':
  unittest.main()
@@ -119,6 +119,22 @@ class TestAssign(unittest.TestCase):
    new = a + old_a
    np.testing.assert_allclose(new.numpy(), 4)

+  def test_assign_changes_alt(self, realize=False):
+    a = Tensor(1).contiguous()
+    if realize: a.realize()
+    b = a.contiguous()    # b returns a new Tensor
+    b.assign(2)
+    b.realize()
+    self.assertNotEqual(a.item(), b.item())
+  # on a realized Tensor contiguous child changes the source
+  @unittest.expectedFailure
+  def test_assign_changes_realized_alt(self): return self.test_assign_changes_alt(realize=True)
+
+  def test_assign_changes_buffer_alt(self):
+    a, b = [Tensor(Tensor(0).contiguous().realize().uop.as_buf()) for _ in range(2)]
+    Tensor.realize(a.contiguous().assign(1), b.contiguous().assign(2))
+    self.assertEqual((a + b).item(), 3)
+
  def test_assign_diamond_cycle(self):
    # NOTE: should *not* raise AssertionError from numpy
    with self.assertRaisesRegex(RuntimeError, "cycle"):
@@ -7,7 +7,7 @@ from tinygrad.helpers import getenv, DEBUG, CI
 from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8, _to_np_dtype, _to_torch_dtype, truncate
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad import Device, Tensor, dtypes
-from hypothesis import assume, given, settings, strategies as strat
+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
 import pytest
@@ -52,8 +52,6 @@ def _test_cast(a:Tensor, target_dtype:DType):
  if target_dtype in dtypes.fp8s: expected = list(map(lambda x: truncate[target_dtype](x), expected))
  _test_op(lambda: a.cast(target_dtype), target_dtype, expected)
 def _test_bitcast(a:Tensor, target_dtype:DType, target=None):
-  if isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) and a.dtype == dtypes.int8 and target_dtype.itemsize != a.dtype.itemsize:
-    raise unittest.SkipTest("shape changing bitcast of int8 broken on PTX")
  expected = torch.tensor(a.tolist(), dtype=_to_torch_storage_type(a.dtype)).view(_to_torch_dtype(target_dtype)).tolist()
  if target_dtype in dtypes.fp8s: expected = list(map(lambda x: fp8_to_float(x, target_dtype), expected))
  _test_op(lambda: a.bitcast(target_dtype), target_dtype, target or expected)
@@ -294,7 +292,6 @@ class TestInt8DType(TestDType):
  def test_int8_to_uint16_negative(self):
    _test_op(lambda: Tensor([-1, -2, -3, -4], dtype=dtypes.int8).cast(dtypes.uint16), dtypes.uint16, [2**16-1, 2**16-2, 2**16-3, 2**16-4])

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "broken in ptx")
  def test_bitcast_alt(self):
    a = Tensor([72, -90, 27, 40, -53, 70, 96, 51], dtype=dtypes.int8).bitcast(dtypes.short)
    self.assertListEqual(a.tolist(), [-22968, 10267, 18123, 13152])
@@ -308,8 +305,6 @@ class TestUint8DType(TestDType):
 class TestBitCast(unittest.TestCase):
  @given(strat.sampled_from(dtype_ints + dtype_floats), strat.sampled_from(dtype_ints + dtype_floats))
  def test_shape_change_bitcast(self, dt1, dt2):
-    # NOTE: this has to be assume to prevent hypothesis from skipping all samples
-    assume(not (isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) and dt1 == dtypes.int8)) # TODO: bitcasting int8 fails in PTX
    data = rand_for_dtype(dt1, 32).reshape(2, 2, 8)
    expected = torch.tensor(data.tolist(), dtype=_to_torch_storage_type(dt1)).view(_to_torch_dtype(dt2))
    if dt2 in dtypes.fp8s:
@@ -139,7 +139,7 @@ class TestImageDType(unittest.TestCase):
      # NOTE: the w1 grad must realize to a seperate kernel
      assert w1.grad.uop.is_realized, f"never realized {w1.grad}"
      self.assertEqual(w1.grad.uop.base.buffer.dtype, dtypes.float32)
-      self.assertEqual(len(sched), 8 if RANGEIFY else 10)
+      self.assertEqual(len(sched), 9 if RANGEIFY else 10)

@unittest.skipUnless(REAL_DEV in IMAGE_SUPPORTED_DEVICES, "Images not supported")
 class TestImageRealization(unittest.TestCase):
@@ -1925,6 +1925,16 @@ class TestSchedule(unittest.TestCase):
    run_schedule(check_schedule(loss, 4))
    np.testing.assert_allclose(loss.item(), 0.878309, atol=1e-5, rtol=1e-6)

+  @expect_rangeify_fails
+  def test_const_folding_alt(self):
+    t = Tensor.full((2,), 1.)
+    lt = (t < 0.)
+    a = Tensor.empty(2).assign(t*lt.where(-1., 0.))
+    b = Tensor.empty(2, dtype=dtypes.bool).assign(lt)
+    Tensor.realize(a, b)
+    self.assertEqual(a.tolist(), [0., 0.])
+    self.assertEqual(b.tolist(), [False, False])
+
  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "Validation error on WebGPU")
  def test_mnist_val(self):
    from tinygrad.nn.datasets import mnist
@@ -15,6 +15,10 @@ class TestTiny(unittest.TestCase):
    out = Tensor([1.,2,3])
    self.assertListEqual(out.tolist(), [1.0, 2.0, 3.0])

+  def test_elu(self):
+    out = Tensor([[1.,2],[3,4]]).sum(axis=1).elu()
+    self.assertListEqual(out.tolist(), [3.0, 7.0])
+
  def test_plus(self):
    out = Tensor([1.,2,3]) + Tensor([4.,5,6])
    self.assertListEqual(out.tolist(), [5.0, 7.0, 9.0])
@@ -417,6 +417,11 @@ class TestUOpGraph(unittest.TestCase):
      uops = to_uops_list([v.bitcast(dt)])
      self.assertEqual(len([x for x in uops if x.op is Ops.BITCAST]), 0, f"dtype = {dt}")

+  def test_sub_with_cast_folds(self):
+    a = Variable("a", 0, 5)
+    uops = to_uops_list([a.cast(dtypes.int)+(-a).cast(dtypes.int)])
+    assert uops == [UOp.const(dtypes.int, 0)]
+
  def test_where_on_gated_load_fold(self):
    ridx0 = UOp.range(100, 0)
    d0 = UOp(Ops.DEFINE_GLOBAL, dtypes.long.ptr(), (), 0)
@@ -607,12 +612,13 @@ class TestUOpGraph(unittest.TestCase):
      with self.assertRaises(RuntimeError): to_uops_list([ld1])

  def test_bounds_with_loaded_bool(self):
-    glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(16), (), 0)
-    glbl1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(8), (), 0)
-    gidx0 = UOp(Ops.SPECIAL, dtypes.index, (UOp.const(dtypes.index, 16),), "gidx0")
-    ld0 = glbl0.index(gidx0).load()
-    ld1 = glbl1.index(gidx0.valid(ld0)).load()
-    with self.assertRaises(RuntimeError): to_uops_list([ld1])
+    with Context(IGNORE_OOB=0):
+      glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(16), (), 0)
+      glbl1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(8), (), 0)
+      gidx0 = UOp(Ops.SPECIAL, dtypes.index, (UOp.const(dtypes.index, 16),), "gidx0")
+      ld0 = glbl0.index(gidx0).load()
+      ld1 = glbl1.index(gidx0.valid(ld0)).load()
+      with self.assertRaises(RuntimeError): to_uops_list([ld1])

  def test_fold_gated_load(self):
    glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
@@ -28,22 +28,6 @@ class TestRewriteMap(unittest.TestCase):
    self.assertIs(sub_map[a+b], e)
    self.assertIs(sub_map[(a+b)*c], f)

-  def test_multistage_substitute(self):
-    a = UOp.variable('a', 0, 10)
-    b = UOp.variable('b', 0, 10)
-    c = UOp.variable('c', 0, 10)
-    d = UOp.variable('d', 0, 10)
-    sub1 = {a+b:c}
-    start = (a+b)*c
-    # stage 1: (a+b)*c -> c*c
-    sub_map1 = graph_rewrite_map(start, _substitute, sub1, bottom_up=True)
-    self.assertIs(sub_map1[(a+b)*c], c*c)
-    # stage 2: c*c -> d
-    sub2 = {c*c:d}
-    sub_map2 = graph_rewrite_map(sub_map1[start], _substitute, sub2, input_map=sub_map1, bottom_up=True)
-    # (a+b)*c -> c*c -> d
-    self.assertIs(sub_map2[(a+b)*c], d)
-
  def test_add_zero(self):
    # Build a small graph: add(0, add(const=0, const=5))
    zero_node = UOp.const(dtypes.index, 0)
@@ -144,11 +128,11 @@ class TestRewriteMap(unittest.TestCase):
      yz_sum_zero  = yz_sum + zero_node   -> rewrites to yz_sum
      yz_neg       = -yz_sum_zero        -> -(y+z)
      yz_dneg      = -yz_neg             -> y+z    (double neg gone)
-      x_plus_yz    = x_var + yz_dneg     -> x + (y+z)
-      double_neg_x = -(-x_plus_yz)       -> x + (y+z)
-      final_expr   = double_neg_x * one_node -> x + (y+z)
+      x_plus_yz    = x_var + yz_dneg     -> (x+y)+z  (add nodes get sorted)
+      double_neg_x = -(-x_plus_yz)       -> (x+y)+z
+      final_expr   = double_neg_x * one_node -> (x+y)+z

-    We expect the final result to be (x + (y+z)).
+    We expect the final result to be ((x+y)+z).
    Each original node should map to the final node that replaces it,
    which might be structurally equivalent but not the same reference.
    """
@@ -163,9 +147,9 @@ class TestRewriteMap(unittest.TestCase):
    yz_sum_zero = yz_sum + zero_node    # (y + z) + 0
    yz_neg = -yz_sum_zero               # -(y+z)
    yz_dneg = -yz_neg                   # -(-(y+z)) -> (y+z)
-    x_plus_yz = x_var + yz_dneg         # x + (y+z)
-    double_neg_x = -(-x_plus_yz)        # neg(neg(x+(y+z))) -> x+(y+z)
-    final_expr = double_neg_x * one_node  # (x+(y+z)) * 1 -> x+(y+z)
+    x_plus_yz = x_var + yz_dneg         # x + (y+z) -> (x+y)+z
+    double_neg_x = -(-x_plus_yz)        # neg(neg(x+(y+z))) -> (x+y)+z
+    final_expr = double_neg_x * one_node  # ((x+y)+z) * 1 -> (x+y)+z

    node_map = graph_rewrite_map(final_expr, symbolic)

@@ -182,14 +166,15 @@ class TestRewriteMap(unittest.TestCase):
    # -(-(y+z)) => (y+z)
    self.assertEqual(node_map[yz_dneg], yz_sum)

-    # x + (y+z) => might get recreated if yz_dneg was changed, so compare to x + yz_sum
-    self.assertEqual(node_map[x_plus_yz], x_var + yz_sum)
+    # x + (y+z) => (x+y)+z
+    expected_xyz = (x_var + y_var) + z_var
+    self.assertEqual(node_map[x_plus_yz], expected_xyz)

-    # -(-(x+(y+z))) => x + (y+z)
-    self.assertEqual(node_map[double_neg_x], x_var + yz_sum)
+    # -(-(x+(y+z))) => (x+y)+z
+    self.assertEqual(node_map[double_neg_x], expected_xyz)

-    # (x+(y+z)) * 1 => x+(y+z)
-    self.assertEqual(node_map[final_expr], x_var + yz_sum)
+    # ((x+y)+z) * 1 => (x+y)+z
+    self.assertEqual(node_map[final_expr], expected_xyz)

    # Unchanged atomic nodes map to themselves
    self.assertEqual(node_map[x_var], x_var)
@@ -60,7 +60,7 @@ class TestValidIdxSimplification(unittest.TestCase):
    load = get_gated_load_uop(gate, idx)
    self.check(load,
      "0",
-      "(((lidx0+(gidx0*4))<19)!=True)")
+      "((((gidx0*4)+lidx0)<19)!=True)")

  def test_simplify_within_valid1(self):
    ridx0 = Range(0, 4)
@@ -184,7 +184,6 @@ class TestValidIdxSimplification(unittest.TestCase):
    print("The expressions are not equivalent.")
    print(s.model())

-  @unittest.expectedFailure  # TODO: improve uop_given_valid
  def test_valid_becomes_const2(self):
    ridx0 = Range(0, 4)
    ridx1 = Range(1, 4)
@@ -304,7 +303,7 @@ class TestImageSimplification(unittest.TestCase):
    idx = ((alu4+1530)%1536, alu1+((idx1+((ridx2+7)//8)+31)//32)+(-2))

    load = get_load_image_uop(shape, valid, idx)
-    self.check(load, None, "((((idx1*48)+(r2*6))+r0)+-6)", "(((idx2*2)+r1)+-1)")
+    self.check(load, None, "((((idx1*48)+r0)+(r2*6))+-6)", "(((idx2*2)+r1)+-1)")

  def test_openpilot_conv2(self):
    # conv in test/external/external_test_valid_remove.py
@@ -325,7 +324,7 @@ class TestImageSimplification(unittest.TestCase):
    idx = ((alu3+765)%768, alu1+((idx1+((ridx2+7)//8)+31)//32)+(-2))
    load = get_load_image_uop(shape, valid, idx)

-    self.check(load, None, "((((idx1*24)+(r2*3))+r0)+-3)", "(((idx2*2)+r1)+-1)")
+    self.check(load, None, "((((idx1*24)+r0)+(r2*3))+-3)", "(((idx2*2)+r1)+-1)")

  def test_openpilot_conv3(self):
    # in openpilot 0.9.7
@@ -347,8 +346,8 @@ class TestImageSimplification(unittest.TestCase):

    self.check(load,
               "((((idx2*2)+r0)<11)&((((idx1*8)+r1)<3)!=True))",
-               "(((idx0+((idx1*512)+(r1*64)))+832)%1024)",
-               "((((idx2*2)+r0)+(((idx1+((r1+5)//8))+1)//2))+-4)")
+               "(((idx0+(idx1*512))+(r1*64))+-192)",
+               "((((idx2*2)+(((idx1+((r1+5)//8))+1)//2))+r0)+-4)")

  def test_simplify1(self):
    # idx has the form (A % m, A // m + k) and valid has (c0 < A) and (A < c1)
@@ -386,16 +385,16 @@ class TestImageSimplification(unittest.TestCase):

    # 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%8)//2)")
+    self.check(load, "(idx0<256)", "((((idx0//32)+((idx0%8)*32))+8)%64)", "((idx0%8)//2)")

    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%8)//2)")
+    self.check(load, "(idx0<256)", "((((idx0//32)+((idx0%8)*32))+16)%64)", "((idx0%8)//2)")

    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%8)//2)")
+    self.check(load, "(idx0<256)", "((((idx0//32)+((idx0%8)*32))+24)%64)", "((idx0%8)//2)")

    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%8)//2)")
+    self.check(load, "(idx0<256)", "(((idx0//32)+((idx0%8)*32))%64)", "((idx0%8)//2)")

  def test_simplify5(self):
    # openpilot 0.9.7, chunk replacement to simplify
@@ -116,6 +116,39 @@ class TestSymbolic(unittest.TestCase):
    self.assertEqual((a*b*3+a*b*b).divide_exact(a*b).simplify(), b+3)
    self.assertEqual((((a*-2)+14)*b).divide_exact(((a*-2)+14)).simplify(), b)

+  def helper_test_factor(self, expr, *factors):
+    factored = expr.factor(*factors)
+    self.check_equal_z3(expr, factored)
+    for fac in factors: self.assertIn(fac, factored.toposort())
+
+  def test_uop_factor(self):
+    a = Variable("a", 0, 8)
+    b = Variable("b", 0, 8)
+    c = Variable("c", 0, 8)
+    self.helper_test_factor((1400*a+2800*b), (a+2*b))
+    self.helper_test_factor((1400*a+2800*b)%9000, (a+2*b))
+    self.helper_test_factor((a+2*b), (a+2*b))
+    self.helper_test_factor((a+c+2*b), (a+2*b))
+    self.helper_test_factor((1400*a+c+2800*b)%9000, (a+2*b))
+    self.helper_test_factor((1399*a+c+2800*b)%9000+1400*a+2800*b, (a+2*b))
+    self.helper_test_factor((1400*a+c+2800*b)%9000+1400*a+2800*b, (a+2*b))
+    # self.assertIsNone((a+c+3*b).factor(a+2*b))
+    # self.assertIsNone((1399*a+c+2800*b).factor(a+2*b))
+
+  def test_uop_multiple_factors(self):
+    a = Variable("a", 0, 8)
+    b = Variable("b", 0, 8)
+    c = Variable("c", 0, 8)
+    d = Variable("d", 0, 8)
+    self.helper_test_factor((1400*a+2800*b+2*c+d), (a+2*b), (2*c+d))
+    self.helper_test_factor((100*a+200*b+5*c), (a+2*b), (5*c))
+    self.helper_test_factor((3*a+6*b+2*c+4*d), (a+2*b), (c+2*d))
+    self.helper_test_factor((7*a+14*b+3*c+6*d), (a+2*b), (3*c+6*d))
+    self.helper_test_factor((10*a+20*b+10*c+30*d), (a+2*b), (c+3*d))
+    self.helper_test_factor((10*c+(10*a+20*b)//3+30*d), (a+2*b), (c+3*d))
+    self.helper_test_factor((10*c+(10*a+20*b)//3+30*d), (a+2*b), (c+3*d))
+    # self.assertIsNone((7*a+14*b+3*c+6*d).factor((a+8*b), (2*c+6*d)))
+
  def test_divide_exact_not(self):
    a = Variable("a", 1, 8)
    b = Variable("b", 1, 8)
@@ -130,13 +163,13 @@ class TestSymbolic(unittest.TestCase):
    a = Variable("a", 0, 8)
    b = Variable("b", 0, 8)
    self.helper_test_variable(a*2+a*3, 0, 8*5, "(a*5)")
-    self.helper_test_variable(b+a*2+a*3, 0, 8*6, "(b+(a*5))")
+    self.helper_test_variable(b+a*2+a*3, 0, 8*6, "((a*5)+b)")

  def test_factorize_no_mul(self):
    a = Variable("a", 0, 8)
    b = Variable("b", 0, 8)
    self.helper_test_variable(a+a*3, 0, 8*4, "(a*4)")
-    self.helper_test_variable((a+b)+a*3, 0, 8*5, "(b+(a*4))")
+    self.helper_test_variable((a+b)+a*3, 0, 8*5, "((a*4)+b)")
    self.helper_test_variable((a*3+b)+b*3, 0, 8*7, "((a*3)+(b*4))")

  def test_neg(self):
@@ -159,8 +192,15 @@ class TestSymbolic(unittest.TestCase):
    b = Variable("b", 0, 8)
    self.helper_test_variable(a+a, 0, 16, "(a*2)")
    self.helper_test_variable((a+b)+b, 0, 24, "(a+(b*2))")
-    self.helper_test_variable((a*3+b)+a, 0, 40, "(b+(a*4))")
-    self.helper_test_variable((a+b)+a*3, 0, 40, "(b+(a*4))")
+    self.helper_test_variable((a*3+b)+a, 0, 40, "((a*4)+b)")
+    self.helper_test_variable((a+b)+a*3, 0, 40, "((a*4)+b)")
+
+  def test_add_self_seperated(self):
+    a = Variable("a", 0, 8)
+    b = Variable("b", 0, 8)
+    c = Variable("c", 0, 8)
+    self.helper_test_variable((a+b)+c+a, 0, 32, "(((a*2)+b)+c)")
+    self.helper_test_variable((a*3+b*2)+c*2+a*5, 0, 96, "(((a*8)+(b*2))+(c*2))")

  def test_sub_self(self):
    a = Variable("a", 0, 8)
@@ -279,7 +319,7 @@ class TestSymbolic(unittest.TestCase):
  def test_mod_congruence_multiple_vars(self):
    self.helper_test_variable((9+9*Variable("x",0,3)+9*Variable("y",0,3))%10, 3, 9, "(((x*-1)+(y*-1))+9)")
    self.helper_test_variable((7+9*Variable("x",0,2)+9*Variable("y",0,2)+Variable("z",0,2))%10, 3, 9,
-                              ("(((z+(x*-1))+(y*-1))+7)", "(((y*-1)+(z+(x*-1)))+7)"))
+                              ("(((z+(x*-1))+(y*-1))+7)", "(((y*-1)+(z+(x*-1)))+7)", "((((x*-1)+(y*-1))+z)+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_div_congruence(self):
@@ -455,7 +495,7 @@ class TestSymbolic(unittest.TestCase):
    ridx1005 = UOp.variable("ridx1005", 0, 2)
    ridx1006 = UOp.variable("ridx1006", 0, 2)
    self.helper_test_variable((lidx1+((gidx1*18)+(ridx1005*18)+(lidx0*162))+(gidx0*2)+(ridx1006*2)+-40)//18, -2, 20,
-      "(((((lidx1+(((gidx1*18)+(ridx1005*18))+(lidx0*162)))+(gidx0*2))+(ridx1006*2))+-40)//18)")
+      "((((((((gidx0*2)+(gidx1*18))+(lidx0*162))+lidx1)+(ridx1005*18))+(ridx1006*2))+-40)//18)")

  def test_add_div(self):
    # careful about the lower bounds and upper bounds
@@ -498,7 +538,7 @@ class TestSymbolic(unittest.TestCase):
    self.helper_test_variable((d1*a*b*d1)//(d1), -1000, 1000, "(a*(b*d1))", test_z3=False)
    self.helper_test_variable((d1*a*d2*b*d1)//(d1*d2),  -1000, 1000, "(a*(b*d1))", test_z3=False)
    self.helper_test_variable((d1*a + b*d1)//(d1), -20, 20, "(a+b)", test_z3=False)
-    self.helper_test_variable((d1*a + b*d1 + c*d1)//(d1), -30, 30, "(c+(a+b))", test_z3=False)
+    self.helper_test_variable((d1*a + b*d1 + c*d1)//(d1), -30, 30, "((a+b)+c)", test_z3=False)
    self.helper_test_variable((3*a*d1 + 9*b*d1)//(3*d1*d2), -40, 40, "(((a+(b*3))//(d2*-1))*-1)", test_z3=False)
    self.helper_test_variable((3*a*d1 + 9*b*d1+3)//(3*d1*d2), -401, 399, "(((((a*d1)+((b*d1)*3))+1)//((d1*d2)*-1))*-1)", test_z3=False)

@@ -508,7 +548,7 @@ class TestSymbolic(unittest.TestCase):
    d = Variable("d", 1, 10)
    self.helper_test_variable((d*a+b)//d, 0, 20, "(a+(b//d))")
    self.helper_test_variable((d*a*20+b)//(5*d), 0, 42, "((a*4)+(b//(d*5)))")
-    self.helper_test_variable((d*a*20+b*d*5+10)//(5*d), 0, 52, "((b+(a*4))+(2//d))")
+    self.helper_test_variable((d*a*20+b*d*5+10)//(5*d), 0, 52, "(((a*4)+b)+(2//d))")

  def test_mod_gcd_factor_neg(self):
    self.helper_test_variable((Variable("a", 0, 10)*-4+4)%8, -4, 4, "((((a*-1)+1)%2)*4)")
@@ -561,9 +601,7 @@ class TestSymbolic(unittest.TestCase):
    lidx2 = Variable("lidx2", 0, 3)
    alu0 = gidx2*640+gidx1*160+(gidx0//5)*2+lidx0*320+lidx1*10
    self.helper_test_variable((alu0+lidx2*2+1)//20, 0, 8192,
-                              ("((((((gidx0//5)+lidx2)//5)+lidx1)//2)+(((gidx2*32)+(gidx1*8))+(lidx0*16)))",
-                               "(((lidx1+((lidx2+(gidx0//5))//5))//2)+((gidx2*32)+((gidx1*8)+(lidx0*16))))",
-                               "((((gidx1*8)+(gidx2*32))+(lidx0*16))+((lidx1+((lidx2+(gidx0//5))//5))//2))"))
+                              ("((((gidx1*8)+(gidx2*32))+(lidx0*16))+((lidx1+((lidx2+(gidx0//5))//5))//2))",))

  def test_sum_div_complex2(self):
    gidx0 = Variable("gidx0", 0, 7)
@@ -641,8 +679,21 @@ class TestSymbolic(unittest.TestCase):
    self.helper_test_variable((gidx//4)*4+gidx%4, 0, 124, "gidx")
    self.helper_test_variable(lidx+gidx%4+(gidx//4)*4, 0, 248, "(gidx+lidx)")
    self.helper_test_variable(lidx+(gidx//4)*4+gidx%4, 0, 248, "(gidx+lidx)")
-    self.helper_test_variable(lidx+(gidx//4)*8+2*(gidx%4), 0, 372, "(lidx+(gidx*2))")
-    self.helper_test_variable(lidx+2*(gidx%4)+(gidx//4)*8, 0, 372, "(lidx+(gidx*2))")
+    self.helper_test_variable(lidx+(gidx//4)*8+2*(gidx%4), 0, 372, "((gidx*2)+lidx)")
+    self.helper_test_variable(lidx+2*(gidx%4)+(gidx//4)*8, 0, 372, "((gidx*2)+lidx)")
+
+  def test_div_mod_recombine_seperated(self):
+    gidx = Variable("gidx", 0, 124)
+    lidx = Variable("lidx", 0, 124)
+    a = Variable("a", 0, 3)
+    b = Variable("b", 0, 3)
+    c = Variable("c", 0, 3)
+    self.helper_test_variable(gidx%4+a+b+c+(gidx//4)*4, 0, 133, "(((a+b)+c)+gidx)")
+    self.helper_test_variable((gidx//4)*4+a+b*10+gidx%4, 0, 157, "((a+(b*10))+gidx)")
+    self.helper_test_variable(lidx+gidx%4+a+b+c//2+(gidx//4)*4, 0, 255, "((((a+b)+gidx)+lidx)+(c//2))")
+    self.helper_test_variable(lidx+(gidx//4)*8+b+c+a*8+2*(gidx%4), 0, 402, "(((((a*8)+b)+c)+(gidx*2))+lidx)")
+    # TODO: need better sorting for this one
+    # self.helper_test_variable(lidx+(gidx//4)*4+a*3+b*3+(c*10)%3+gidx%4, , , "")

  def test_div_mod_recombine_folded_mod(self):
    a = Variable("a", 0, 2)
@@ -1019,6 +1070,7 @@ class TestSymbolicRealWorld(unittest.TestCase):
      ("((((((((((lidx5+1)//16)*802816)+(((lidx5+1)%16)*49))+(gidx0*3211264))+(gidx1*784))+(gidx2*8))+(lidx4*100352))+lidx3)+2207744)",
       '((lidx3+((((((((lidx5+1)//16)*802816)+(((lidx5+1)%16)*49))+(gidx0*3211264))+(gidx1*784))+(gidx2*8))+(lidx4*100352)))+2207744)',
       '((lidx3+((lidx4*100352)+((gidx2*8)+((gidx1*784)+((gidx0*3211264)+((((lidx5+1)//16)*802816)+(((lidx5+1)%16)*49)))))))+2207744)',
+       '((((((((gidx0*3211264)+(gidx1*784))+(gidx2*8))+lidx3)+(lidx4*100352))+(((lidx5+1)//16)*802816))+(((lidx5+1)%16)*49))+2207744)',
      ))

 class TestBounds(unittest.TestCase):
@@ -42,7 +42,7 @@ class TestWinograd(unittest.TestCase):
    out = Tensor.conv2d(x,w, padding=1)
    out.mean().backward()
    backward_schedule = Tensor.schedule(x.grad, w.grad)
-    self.assertEqual(len(backward_schedule), 3 if RANGEIFY else 9)
+    self.assertEqual(len(backward_schedule), 4 if RANGEIFY else 9)

  def test_counters(self):
    IC, OC, X, Y = 4,4,9,9
@@ -135,7 +135,7 @@ class Transformer:
    x = self.token_embd(tokens)                           # (B, T, D)
    for block in self.blk: x = block(x, start_pos)
    # TODO: add temperature
-    return self.output(self.output_norm(x))[:, -1, :].softmax(-1).argmax(-1, keepdim=True)
+    return self.output(self.output_norm(x))[:, -1, :].softmax(-1, dtype="float").argmax(-1, keepdim=True)

  def __call__(self, tokens:Tensor, start_pos:int|UOp=0) -> Tensor:
    return (self.forward_jit if getenv("JIT", 1) and tokens.shape[1] == 1 and isinstance(start_pos, UOp) else self.forward)(tokens, start_pos)
@@ -10,7 +10,7 @@ from tinygrad.renderer import Renderer
 from tinygrad.codegen.lowerer import pm_lowerer, get_index
 from tinygrad.codegen.quantize import pm_quant
 from tinygrad.codegen.gpudims import pm_add_gpudims
-from tinygrad.uop.symbolic import sym, symbolic_simple, gep_pushing
+from tinygrad.uop.symbolic import sym, symbolic_simple, gep_pushing, symbolic
 from tinygrad.uop.decompositions import get_late_rewrite_patterns
 from tinygrad.codegen.late.expander import migrate_indexing, expander, pm_pre_expander, pm_group_for_reduce
 from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
@@ -104,6 +104,7 @@ def _get_rewrites_for_renderer(opts:Renderer, optimize:bool, linearizer:bool, _Q

  # lower the index dtype to a concrete int
  ret.append(RewriteStep(pm_lower_index_dtype+load_store_indexing, lambda _: opts.device, name="lower all index dtypes"))
+  ret.append(RewriteStep(symbolic, name="post index symbolic"))

  # optional pre matcher
  if opts.pre_matcher is not None: ret.append(RewriteStep(opts.pre_matcher, name="pre_matcher"))
@@ -148,6 +148,8 @@ CPU_COUNT = ContextVar("CPU_COUNT", max(1, (os.cpu_count() or 1) // (4 if ARCH_X
 CPU_LLVM, AMD_LLVM = ContextVar("CPU_LLVM", 0), ContextVar("AMD_LLVM", 1)
 VIZ = PROFILE = ContextVar("VIZ", 0)
 SPEC = ContextVar("SPEC", 0)
+# TODO: disable by default due to speed
+IGNORE_OOB = ContextVar("IGNORE_OOB", 1)

@dataclass(frozen=True)
 class Metadata:
@@ -102,7 +102,7 @@ string_rewrite = PatternMatcher([
  (UPat(Ops.CAST, name="x", dtype=dtypes.bool, src=(UPat.var("a"),)),
   lambda ctx, x, a: f"setp.ne.b{ctx.types[a.dtype][1:]} {ctx.r[x]}, {ctx.r[a]}, {render_val(0, a.dtype)};"),
  (UPat(Ops.CAST, name="x", src=(UPat.var("a"),)),
-   lambda ctx, x, a: f"cvt{modifier(x.dtype, a.dtype)}.{ctx.types[x.dtype]}.{ctx.types[a.dtype]} {ctx.r[x]}, {ctx.r[a]};"),
+   lambda ctx, x, a: f"cvt{modifier(x.dtype, a.dtype)}.{ctx.cast_types[x.dtype]}.{ctx.cast_types[a.dtype]} {ctx.r[x]}, {ctx.r[a]};"),
  (UPat(Ops.LOAD, name="x", src=(UPat.var('loc'), UPat(name='alt'), UPat(name="gate", op=GroupOp.ALU)), allow_any_len=True),
    lambda ctx, x, loc, alt, gate: flatten([
    [f"mov.{ctx.mem_types[x.dtype.scalar()]} {v}, {render_val(0, x.dtype.scalar())};" for v in ctx.r[x]],
@@ -146,12 +146,12 @@ class PTXRenderer(Renderer):
 .address_size 64
 .visible .entry"""
  barrier = "bar.sync\t0;"
-  # HACK: Use s16 and u16 for int8 and uint8 buffers. This can be wrong in cast.
  types: dict[DType, str] = { dtypes.int8: "s16", dtypes.int16: "s16", dtypes.int32: "s32", dtypes.int64: "s64",
                              dtypes.uint8: "u16", dtypes.uint16: "u16", dtypes.uint32: "u32", dtypes.uint64: "u64",
                              dtypes.float16: "f16", dtypes.float32: "f32", dtypes.float64: "f64", dtypes.bool: "pred" }

  mem_types: dict[DType, str] = {**types, dtypes.int8: "s8", dtypes.uint8: "u8", dtypes.bool: "u8", dtypes.float16: "b16"}
+  cast_types: dict[DType, str] = {**types, dtypes.int8: "s8", dtypes.uint8: "u8"}

  def render_kernel(self, kernel, function_name, bufs, regs, uops) -> str:
    def fmt(line): return line if line[0]=="$" else "\t" + line.replace(" ", "\t" if len(line.split(" ")[0]) > 7 else "\t\t", 1)
@@ -176,7 +176,7 @@ class RemoteHandler:
    self.sessions: defaultdict[SessionKey, RemoteSession] = defaultdict(RemoteSession)

    try: self.ib_ctx: IBCtx|None = IBCtx(getenv("IB_DEV", 0))
-    except (IndexError, AttributeError): self.ib_ctx = None
+    except (RuntimeError, IndexError, AttributeError): self.ib_ctx = None
    self.ib_lock = asyncio.Lock()
    self.ib_conns: dict[str, IBConn|None] = {}
    self.iova_cache: dict[tuple[SessionKey, int], tuple[int, int, int]] = {}
@@ -10,7 +10,7 @@ DEFAULT_PORT, DEFAULT_GID = getenv("DEFAULT_PORT", 1), getenv("DEFAULT_GID", 3)
 IOVA_ALIGN = resource.getpagesize()

 def checkz(x, ret=None):
-  assert x == 0, f'{x} != 0 (errno {ctypes.get_errno()})'
+  if x != 0: raise RuntimeError(f'{x} != 0 (errno {ctypes.get_errno()})')
  return ret

@dataclass(frozen=True)
@@ -81,8 +81,8 @@ earliest_rewrites = PatternMatcher([
  # copy only to different device
  (UPat(Ops.COPY, src=(UPat.var("x"), UPat()), name="copy"), lambda x,copy: x.f(Ops.NOOP, tag=copy.tag) if x.device == copy.device else None),

-  # contiguous/buffer/copy/assign is already contiguous
-  #(UPat(Ops.CONTIGUOUS, name="root", src=(UPat((Ops.CONTIGUOUS, Ops.BUFFER, Ops.COPY, Ops.ASSIGN)),)), lambda root: root.src[0]),
+  # contiguous buffer is buffer, this is for *correctness* of assign, not just speed
+  (UPat(Ops.CONTIGUOUS, name="root", src=(UPat(Ops.BUFFER),)), lambda root: root.src[0].forced_reshape(root.shape).rtag(root.tag)),
 ])

 # *****************
@@ -270,12 +270,11 @@ def map_reduce(ctx:RangeifyContext, idx:UOp, red:UOp):

 def index_child(ctx:RangeifyContext, c:UOp, x:UOp, idx:UOp):
  if c not in ctx.seen_children: ctx.seen_children[c] = {}
+  ctx.seen_children[c][x.arg[0]] = idx
  # wait here until we have seen all the children
  if len(ctx.seen_children[c]) != x.arg[1]:
    ctx.progress += 1
    if ctx.progress > 10000: raise RuntimeError("children not making progress")
-    # NOTE: we mark this here
-    ctx.seen_children[c][x.arg[0]] = idx
    raise RewriteNotReady
  ctx.progress = 0

@@ -431,7 +430,7 @@ def remove_bufferize(src:UOp, buf:UOp, idx:UOp):

    # const reduce is okay
    # TODO: move the reduce folder to before this to prevent the need for this
-    def okay_reduce(x:UOp): return all(y.op not in {Ops.BUFFER, Ops.COPY} for y in x.sparents)
+    def okay_reduce(x:UOp): return all(y.op not in {Ops.BUFFER, Ops.BUFFERIZE, Ops.COPY} for y in x.sparents)

    # always run this list of ops
    if any(x.op is Ops.REDUCE and not okay_reduce(x) for x in ran): return None
@@ -439,7 +438,8 @@ def remove_bufferize(src:UOp, buf:UOp, idx:UOp):
  # if it makes it here, the bufferize is removed
  # this is the ranges replaced
  # NOTE: if buf src is a const, we don't replace it
-  return src.substitute({k:v for k,v in zip(buf.src[1:], idx.src[1:]) if k.op is not Ops.CONST})
+  replaces = flatten([(k,v) for k,v in zip(buf.src[1:], idx.src[1:]) if k.op is not Ops.CONST])
+  return UOp(Ops.SUBSTITUTE, src=(src, UOp(Ops.NOOP, src=tuple(replaces[0::2])), UOp(Ops.NOOP, src=tuple(replaces[1::2]))))

 def pre_bufferize(b:UOp, x:UOp, copy:UOp):
  nb = b.replace(src=(b.src[0].contiguous(),)+b.src[1:])
@@ -714,6 +714,24 @@ replace_contiguous = PatternMatcher([
  (UPat(GroupOp.ALU, name="alu"), lambda ctx,alu: alu.replace(src=new_src) if (new_src:=tuple(ctx.get(s, s) for s in alu.src)) != alu.src else None),
 ])

+def do_sub_recurse(s:UOp):
+  x,keys,values = s.src[0], s.src[1].src, s.src[2].src
+  # SUBSTITUTE applied to SUBSTITUTE runs the child SUB on the parents. though this is probably wrong in the generic case
+  if x.op is Ops.SUBSTITUTE:
+    sub_k = UOp(Ops.SUBSTITUTE, src=(x.src[1],)+s.src[1:])
+    sub_v = UOp(Ops.SUBSTITUTE, src=(x.src[2],)+s.src[1:])
+    return UOp(Ops.SUBSTITUTE, src=(x.src[0], sub_k, sub_v))
+  # here we actually do the SUBSTITUTE
+  if x in keys: return values[keys.index(x)]
+  # we filter any keys that aren't in parents. this keeps the algorithm O(output graph size)
+  new_kv = {k:v for k,v in zip(keys,values) if k in x.sparents}
+  # if there's no SUBSTITUTEs left, we can just return x
+  if len(new_kv) == 0: return x
+  # then we add SUBSTITUTE to all parents
+  uop_keys, uop_values = UOp(Ops.NOOP, src=tuple(new_kv.keys())), UOp(Ops.NOOP, src=tuple(new_kv.values()))
+  return x.replace(src=tuple([UOp(Ops.SUBSTITUTE, src=(y,uop_keys,uop_values)) for y in x.src]))
+pm_substitute_recurse = PatternMatcher([(UPat(Ops.SUBSTITUTE, src=(UPat(), UPat(Ops.NOOP), UPat(Ops.NOOP)), name="s"), do_sub_recurse)])
+
@track_rewrites(lambda _,ret: f"Schedule {pluralize('Kernel', len([u for u in UOp.sink(*ret.values()).toposort() if u.op is Ops.KERNEL]))}", True)
 def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
  uop_list: list[UOp] = []
@@ -731,7 +749,7 @@ def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
  tsink = graph_rewrite(tsink, pm_rangeify, ctx=(rangeify_ctx:=RangeifyContext()), bottom_up=True, name="rangeify")
  # NOTE: sym (vs symbolic_simple) breaks things here because ranges with len 1 aren't handled right
  tsink = graph_rewrite(tsink, symbolic_simple+pm_reduce_unparented, name="symbolic")  # this supports const folding
-  tsink = graph_rewrite(tsink, pm_cleanups, bottom_up=True, name="remove costly buffers")
+  tsink = graph_rewrite(tsink, pm_cleanups+pm_substitute_recurse, bottom_up=True, name="remove costly buffers")
  tsink = graph_rewrite(tsink, pm_limit_bufs, ctx=rangeify_ctx, name="limit buffers")

  # rebuild the sink with all the BUFFERIZEs with tags, this is what's ending up in the tensor graph
@@ -10,6 +10,7 @@ class FastEnum(IntEnum):
 class Ops(FastEnum):
  # uops that aren't rendered
  NOOP = auto(); SINK = auto(); UNIQUE = auto(); DEVICE = auto(); KERNEL = auto(); PRECAST = auto(); REWRITE_ERROR = auto()  # noqa: E702
+  SENTINEL = auto()

  # track children
  CHILD = auto(); CHILDREN = auto() # noqa: E702
@@ -19,6 +20,7 @@ class Ops(FastEnum):

  # create buffer
  BUFFERIZE = auto()
+  SUBSTITUTE = auto()

  # ops that adjust the behavior of the scheduler
  CONTIGUOUS = auto(); CONTIGUOUS_BACKWARD = auto(); DETACH = auto(); FUSE = auto() # noqa: E702
@@ -8,7 +8,7 @@ from tinygrad.uop.mathtraits import MathTrait
 from tinygrad.dtype import ConstType, ImageDType, dtypes, DType, truncate, PtrDType, least_upper_dtype, Invalid, InvalidType
 from tinygrad.helpers import ContextVar, all_int, prod, getenv, all_same, Context, partition, temp, unwrap, T, argfix, Metadata, flatten, TRACEMETA
 from tinygrad.helpers import PICKLE_BUFFERS, PROFILE, dedup, cdiv, cmod, diskcache_put, to_function_name, cpu_profile, TracingKey, RANGEIFY, VIZ, SPEC
-from tinygrad.helpers import strip_parens
+from tinygrad.helpers import strip_parens, make_tuple
 if TYPE_CHECKING:
  from tinygrad.shape.shapetracker import ShapeTracker
  from tinygrad.device import Buffer, MultiBuffer
@@ -150,6 +150,11 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
  def tuplize(self:UOp) -> tuple:
    return (self.op.value, self.arg, self.dtype,)+tuple([x.tuplize for x in self.src])

+  @functools.cached_property
+  def order_add(self:UOp) -> tuple:
+    if self.op is Ops.MUL and self.src[1].op in (Ops.CONST, Ops.VCONST): return (self.src[0].tuplize, make_tuple(self.src[1].arg, 1))
+    return (self.tuplize, (0,))
+
  @property
  def ptrdtype(self) -> PtrDType:
    if not isinstance(self.dtype, PtrDType): raise RuntimeError("ptrdtype called on UOp without PtrDType")
@@ -241,9 +246,9 @@ class UOp(MathTrait, metaclass=UOpMetaClass):

  def simplify(self, tracked=False):
    # late import!
-    from tinygrad.uop.symbolic import symbolic
+    from tinygrad.uop.symbolic import symbolic_flat
    with Context(TRACK_MATCH_STATS=0 if not tracked else TRACK_MATCH_STATS.value):
-      return graph_rewrite(self, symbolic, name="simplify")
+      return graph_rewrite(self, symbolic_flat, name="simplify")
  def ssimplify(self) -> UOp|ConstType: return ret.arg if (ret:=self.simplify()).op is Ops.CONST else ret
  def _eval(self, dtype, expected_type:Type[T]) -> T:
    assert self.dtype in dtype, f"eval with wrong dtype {self}"
@@ -568,6 +573,32 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
      if (d0:=self.src[0].divides(v)) is not None: return d0 * self.src[1]
      if (d1:=self.src[1].divides(v)) is not None: return self.src[0] * d1
    return None # generic None if we aren't sure
+  def factor(self, *factors: UOp) -> UOp:
+    # factor out expr from self if possible, might return self
+    # (1400*a + 2800*b + c).factor(a+2*b) -> 1400*(a+2*b) + c
+    if self.dtype in dtypes.floats: return self
+    if self.op is Ops.ADD:
+      factored = []
+      # dict of {term: const_factor}, i.e. {a: 1, b: 2}
+      remainders = dict([(u.divides(f:=u.const_factor()).simplify(),f) for u in self.split_uop(Ops.ADD)])
+      for fac in factors:
+        if fac.dtype not in (dtypes.index,)+dtypes.ints: continue
+        fac_terms = dict((u.divides(f:=u.const_factor()).simplify(),f) for u in fac.split_uop(Ops.ADD))
+        factored_terms  = {k:v for k,v in remainders.items() if k in fac_terms}
+        new_remainders  = {k:v for k,v in remainders.items() if k not in fac_terms}
+
+        if any(u not in factored_terms for u in fac_terms) or any(factored_terms[u]%fac_terms[u]!=0 for u in fac_terms) or not \
+          all_same(mul:=[factored_terms[u]//fac_terms[u] for u in fac_terms]):
+          continue
+
+        remainders = new_remainders
+        factored.append(fac*mul[0])
+      if not factored: return self
+      start = functools.reduce(operator.add, factored)
+      return sum([k.factor(*factors)*v for k,v in remainders.items()], start=start)
+
+    if self.op not in GroupOp.ALU|{Ops.VECTORIZE}: return self
+    return self.replace(src=tuple(s.factor(*factors) for s in self.src))
  def pop_const(self, op=Ops.ADD) -> tuple[UOp, ConstType]:
    return (self.src[0], self.src[1].arg) if self.op is op and self.src[1].op is Ops.CONST else (self, identity_element(op, self.dtype))
  @staticmethod
@@ -992,6 +1023,7 @@ if TRACK_MATCH_STATS or PROFILE:

 # *** simple graph rewrite engine ***

+SENTINEL = UOp(Ops.SENTINEL)
 class RewriteNotReady(Exception): pass
 class BottomUpGate(Exception): pass
 class RewriteContext:
@@ -1004,45 +1036,58 @@ class RewriteContext:
    self.replace: dict[UOp, UOp] = {}

  def cached_pm_rewrite(self, x:UOp):
-    if (ret:=self.pm_cache.get(x,False)) is not False: return ret
+    if (ret:=self.pm_cache.get(x,SENTINEL)) is not SENTINEL: return ret
    ret = self.pm_cache[x] = cast(PatternMatcher, self.pm).rewrite(x, self.ctx)
    return ret

  def cached_bpm_rewrite(self, x:UOp):
-    if (ret:=self.bpm_cache.get(x,False)) is not False: return ret
+    if (ret:=self.bpm_cache.get(x,SENTINEL)) is not SENTINEL: return ret
    ret = self.bpm_cache[x] = cast(PatternMatcher, self.bpm).rewrite(x, self.ctx)
    return ret

  def unified_rewrite(self, root:UOp) -> UOp:
    stack: collections.deque[tuple[UOp, int, UOp]] = collections.deque([(root, 0, root)])
    on_stack = {root}  # all UOps either on the stack or in self.replace, i.e. dont have to be placed again
+    REWRITE_STACK_LIMIT = getenv("REWRITE_STACK_LIMIT", 250000)
    while stack:
-      if len(stack) > getenv("REWRITE_STACK_LIMIT", 250000): raise RuntimeError("infinite loop in graph_rewrite (stack too big)")
+      if len(stack) > REWRITE_STACK_LIMIT: raise RuntimeError("infinite loop in graph_rewrite (stack too big)")
      n, stage, new_n = stack.pop()
      if n in self.replace: continue  # skip any nodes we have seen
-      try:
-        if stage == 0:
+      if stage == 0:
+        # if bottom up, we rewrite this node early. in both cases, we add its parents to the stack
+        if self.bpm is not None:
+          # apply rewrite rules until a fixed point is reached. may return `uop` itself if PatternMatcher doesn't match
+          test_n: UOp|None = n
+          seen = set()
          try:
-            # if bottom up, we rewrite this node early. in both cases, we add its parents to the stack
-            if self.bpm is not None:
-              # apply rewrite rules until a fixed point is reached. may return `uop` itself if PatternMatcher doesn't match
-              test_n: UOp|None = n
-              seen = set()
-              while test_n is not None:
-                if test_n in seen: raise RuntimeError("infinite loop in fixed_point_rewrite")
-                seen.add(test_n)
-                new_n, test_n = test_n, self.cached_bpm_rewrite(test_n)
-            stack.append((n, 1, new_n))
-            for x in reversed(new_n.src):
-              if x in on_stack: continue
-              stack.append((x, 0, x))
-              on_stack.add(x)
-          # if the bpm matching raised a gate, we are done with this node and dont continue down the srcs
-          except BottomUpGate: self.replace[n] = new_n
-        elif stage == 1:
-          try: new_src = tuple([self.replace[x] for x in new_n.src])
-          except KeyError: raise RewriteNotReady
-          if new_src == new_n.src:
+            while test_n is not None:
+              if test_n in seen: raise RuntimeError("infinite loop in fixed_point_rewrite")
+              seen.add(test_n)
+              new_n, test_n = test_n, self.cached_bpm_rewrite(test_n)
+          except RewriteNotReady:
+            # try the full thing again later
+            stack.appendleft((n, 0, n))
+            continue
+          except BottomUpGate:
+            # if the bpm matching raised a gate, we are done with this node and dont continue down the srcs
+            self.replace[n] = new_n
+            continue
+        stack.append((n, 1, new_n))
+        for x in reversed(new_n.src):
+          if x in on_stack: continue
+          stack.append((x, 0, x))
+          on_stack.add(x)
+      elif stage == 1:
+        tmp = []
+        for x in new_n.src:
+          if (rx:=self.replace.get(x, SENTINEL)) is SENTINEL:
+            # if some new sources aren't ready, we try this again later
+            stack.appendleft((n, 1, new_n))
+            break
+          tmp.append(rx)
+        else:
+          # in stage 1, once all srcs are rewritten, rebuild (if changed) or run top-down rewrite
+          if (new_src:=tuple(tmp)) == new_n.src:
            # if top down, do the rewrite. if no rewrite or bottom up, we are done rewriting this node so we add it to the dict
            if self.pm is None or (new_src_n:=self.cached_pm_rewrite(new_n)) is None:
              self.replace[n] = new_n
@@ -1053,13 +1098,14 @@ class RewriteContext:
          # trigger a rewrite of new_src_n, then after that rewrite is done, link it back to n
          stack.append((n, 2, new_src_n))
          stack.append((new_src_n, 0, new_src_n))
+      else:
+        # in stage 2, we link the result of new_n to the result of n
+        if (replaced_new_n:=self.replace.get(new_n, SENTINEL)) is SENTINEL:
+          # not ready, try the link later
+          stack.appendleft((n, 2, new_n))
        else:
-          # in stage 2, we link the result of new_n to the result of n
-          try: self.replace[n] = self.replace[new_n]
-          except KeyError: raise RewriteNotReady
-      except RewriteNotReady:
-        # retry this later
-        stack.appendleft((n, stage, new_n))
+          # otherwise we are done
+          self.replace[n] = replaced_new_n
    return self.replace[root]

@track_matches
@@ -1,7 +1,7 @@
 from typing import cast, Callable
 from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, print_uops, python_alu, graph_rewrite, AxisType
 from tinygrad.dtype import DType, ImageDType, dtypes, PtrDType, AddrSpace, Invalid
-from tinygrad.helpers import all_same, prod, DEBUG, ContextVar, Context, cpu_profile, RANGEIFY
+from tinygrad.helpers import all_same, prod, DEBUG, IGNORE_OOB, Context, cpu_profile, RANGEIFY
 from tinygrad.shape.shapetracker import ShapeTracker
 try:
  import z3
@@ -55,9 +55,6 @@ try:
  z3_imported = True
 except (ImportError, AttributeError): z3_imported = False

-# if you have z3 installed, by default we check the bounds
-IGNORE_OOB = ContextVar("IGNORE_OOB", int(not z3_imported))
-
 buffer_spec = PatternMatcher([
  (UPat(Ops.UNIQUE, dtypes.void, ()), lambda: True),
  (UPat(Ops.DEVICE, dtypes.void, (), name="d"), lambda d:
@@ -274,10 +274,16 @@ gep_pushing = PatternMatcher([
  (UPat(Ops.WMMA, name="wmma").f(Ops.GEP, name="gep"), gep_through_wmma),
 ])

+def chain_insert(chain, b, op):
+  if chain.op is not op or b.order_add > chain.src[1].order_add: return chain.alu(op, b)
+  return chain_insert(chain.src[0], b, op).alu(op, chain.src[1])
+
 commutative = PatternMatcher([
  # ** COMMUTATIVE flipping (only for index) **
  # NOTE: this can break merging vector math by only flipping some of them
-  (UPat(GroupOp.Commutative, dtype=dtypes.index, name='x'), lambda x: x.replace(src=x.src[::-1]) if x.src[1].tuplize < x.src[0].tuplize else None),
+  (UPat(GroupOp.Commutative-{Ops.ADD}, dtype=dtypes.index, name='x'), lambda x:
+    x.replace(src=x.src[::-1]) if x.src[1].tuplize < x.src[0].tuplize else None),
+  (UPat(Ops.ADD, dtype=dtypes.index, name="x"), lambda x: functools.reduce(operator.add, sorted(x.split_uop(Ops.ADD), key=lambda u: u.order_add)))
 ])

 symbolic = symbolic_simple+commutative+PatternMatcher([
@@ -373,7 +379,7 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
 ])+gep_pushing

 symbolic_flat = symbolic+PatternMatcher([
-  # ** combine terms (opinionated) **
+  # ** combine terms (opinionated), can make it harder to substitute valids **
  (-1 * (UPat.var("x") + UPat.var("y")), lambda x,y: (-x)+(-y)),  # -(x+y) -> -x + -y
  # (x+y)*c -> x*c+y*c. only for int, float has inf*0=nan issue
  ((UPat.var("x", dtypes.index) + UPat.var("y")) * UPat.cvar("c"), lambda x,y,c: x*c+y*c),
@@ -405,10 +411,13 @@ def uop_given_valid(valid:UOp, uop:UOp) -> UOp|None:
  # don't simplify any other gates, can lead to OOB, we substitute them back later
  uop = uop.substitute((load_subs:={u: UOp(Ops.NOOP, arg=u) for u in uop.toposort() if u.op is Ops.INDEX}))

+  all_candidates = []
  # simplify uop given that valid is True
-  for expr,v in bounds.items():
+  for i, (expr,v) in enumerate(bounds.items()):
    v0, v1 = (expr.vmin if v[0] is None else v[0], expr.vmax if v[1] is None else v[1])
    expr = expr.substitute(load_subs)  # make sure expr appears in same form in the uop
+    # if the expr is an add we try and factorize so its more likely to substitute
+    if expr.op is Ops.ADD: uop = uop.factor(expr)
    # some expr has lower bound > upper bound -> valid is an empty set and we return None
    if v0 > v1: return None
    # whole node became a const
@@ -421,7 +430,9 @@ def uop_given_valid(valid:UOp, uop:UOp) -> UOp|None:
      # if the constraint is a simplex: X0 + X1 + ... > 0, we can check if all Xi > 0 simplify into the same output
      candidates.append([(Xi, UOp.variable("fake", 1, Xi.vmax, Xi.dtype)) for Xi in expr.split_uop(Ops.ADD)])
    # try checking the whole clause
-    if expr in uop.toposort(): candidates.append([(expr, UOp.variable("fake", v0, v1, expr.dtype))])
+    if expr in uop.toposort():
+      candidates.append([tup:=(expr, UOp.variable(f"fake{i}", v0, v1, expr.dtype))])
+      all_candidates.append(tup)

    for candidate in candidates:
      # if every branch in candidate gives the same simplified uop, we can rewrite the uop
@@ -431,6 +442,9 @@ def uop_given_valid(valid:UOp, uop:UOp) -> UOp|None:
        if all_same([uops.src[1] for uops in newuops]): uop = uop.replace(src=(uop.src[0], newuops[0].src[1]))
      elif all_same(newuops): uop = newuops[0]

+  uop = uop.factor(*(e[0] for e in all_candidates))
+  uop = uop.substitute(sub_dict:=dict(all_candidates)).simplify().substitute({newX:X for X,newX in sub_dict.items()}).simplify()
+
  # put the loads back in
  uop = uop.substitute({v:k for k,v in load_subs.items()})
  return uop
@@ -20,7 +20,8 @@ uops_colors = {Ops.LOAD: "#ffc0c0", Ops.STORE: "#87CEEB", Ops.CONST: "#e0e0e0",
               **{x:"#D8F9E4" for x in GroupOp.Movement}, **{x:"#ffffc0" for x in GroupOp.ALU}, Ops.THREEFRY:"#ffff80", Ops.BUFFER_VIEW: "#E5EAFF",
               Ops.BLOCK: "#C4A484", Ops.BLOCKEND: "#C4A4A4", Ops.BUFFER: "#B0BDFF", Ops.COPY: "#a040a0", Ops.FUSE: "#FFa500",
               Ops.ALLREDUCE: "#ff40a0", Ops.MSELECT: "#d040a0", Ops.MSTACK: "#d040a0", Ops.CONTIGUOUS: "#FFC14D", Ops.REALIZE: "#C1C14D",
-               Ops.CHILDREN: "#80ffc0", Ops.CHILD: "#80fff0", Ops.BUFFERIZE: "#FF991C", Ops.REWRITE_ERROR: "#ff2e2e"}
+               Ops.CHILDREN: "#80ffc0", Ops.CHILD: "#80fff0", Ops.BUFFERIZE: "#FF991C", Ops.REWRITE_ERROR: "#ff2e2e",
+               Ops.SUBSTITUTE: "#ffff00"}

 # VIZ API