diff --git a/.github/workflows/benchmark.yml b/.github/workflows/benchmark.yml
index 00f0a0e5ec..2c3ac71e68 100644
--- a/.github/workflows/benchmark.yml
+++ b/.github/workflows/benchmark.yml
@@ -692,8 +692,9 @@ jobs:
       run: AMD=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
     - name: Run full CIFAR training w 1 GPU
       run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
+    # TODO: enable
+    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
       run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee am_train_bert_one_gpu.txt
diff --git a/examples/mlperf/dataloader.py b/examples/mlperf/dataloader.py
index 121698c97b..c82c42410d 100644
--- a/examples/mlperf/dataloader.py
+++ b/examples/mlperf/dataloader.py
@@ -564,10 +564,13 @@ class GPTDataset:
     # check for cache
     cache_hash = hashlib.sha256(f"{samples}:{seqlen}:{seed}:{shuffle}".encode()).hexdigest()
     cache_path = base_path.with_name(f"{base_path.name}.{cache_hash}.index_cache")
+    print(f"try loading GPTDataset from {cache_path}...")
     if cache_path.exists():
+      print("cache found, loading...")
       with open(cache_path, "rb") as f:
         self.doc_idx, self.sample_idx, self.shuffle_idx = pickle.load(f)
     else:
+      print("cache not found, building index...")
       self.doc_idx = self._build_doc_idx()
       self.sample_idx = self._build_sample_idx()
       self.shuffle_idx = self._build_shuffle_idx()
@@ -692,10 +695,47 @@ class BlendedGPTDataset:
 
     self.datasets = [GPTDataset(path, samples_per_blend[i], seqlen, seed + i, shuffle) for i,path in enumerate(paths)]
 
+    # check for cache
+    cache_hash = hashlib.sha256(f"{samples}:{seqlen}:{seed}:{shuffle}".encode()).hexdigest()
+    cache_path = paths[0].with_name(f"{paths[0].name}.{cache_hash}.blend_cache")
+    print(f"try loading BlendedGPTDataset from {cache_path}...")
+    if cache_path.exists():
+      print("cache found, loading...")
+      with open(cache_path, "rb") as f:
+        self.dataset_idx, self.dataset_sample_idx = pickle.load(f)
+    else:
+      print("cache not found, building index...")
+      self.dataset_idx, self.dataset_sample_idx = self._build_blend_idx()
+      # save cache
+      with open(cache_path, "wb") as f:
+        pickle.dump((self.dataset_idx, self.dataset_sample_idx), f)
+
   def get(self, idx:int):
-    tokens = self.datasets[0][idx]
+    tokens = self.datasets[self.dataset_idx[idx]][self.dataset_sample_idx[idx]]
     return tokens
 
+  def _build_blend_idx(self):
+    dataset_idx = np.zeros(self.samples, dtype=np.int16)
+    dataset_sample_idx = np.zeros(self.samples, dtype=np.int64)
+
+    unspent_datasets = set(range(len(self.datasets)))
+    dataset_sample_counts = [0] * len(self.datasets)
+
+    for i in tqdm(range(self.samples)):
+      error_argmax, error_max = 0, 0.0
+      for di in unspent_datasets:
+        error = self.weights[di] * max(i, 1) - dataset_sample_counts[di]
+        if error > error_max:
+          error_max = error
+          error_argmax = di
+
+      dataset_idx[i] = error_argmax
+      dataset_sample_idx[i] = dataset_sample_counts[error_argmax]
+
+      dataset_sample_counts[error_argmax] += 1
+
+    return dataset_idx, dataset_sample_idx
+
 def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
   if val:
     dataset = BlendedGPTDataset([
diff --git a/test/test_schedule.py b/test/test_schedule.py
index 87530de80e..59aba3ee74 100644
--- a/test/test_schedule.py
+++ b/test/test_schedule.py
@@ -1727,7 +1727,8 @@ class TestIndexing(unittest.TestCase):
       s = Tensor.schedule(*lst)
       lowered = [x[1] for x in lower_schedule(s.copy())]
       kernels = [ei for ei in list(lowered) if isinstance(ei.prg, CompiledRunner)]
-      if FUSE_ARANGE: self.assertEqual(len(kernels), cnt)
+      if FUSE_ARANGE and len(kernels) != cnt:
+        raise KernelCountException(f"{len(kernels)} != {cnt}")
       for ei in lowered: ei.run(do_update_stats=True)
     return s
 
@@ -1741,7 +1742,7 @@ class TestIndexing(unittest.TestCase):
   def test_simple_indexing_alt(self):
     X = Tensor.arange(16).reshape(4, 4)
     xt = X[[1, 2], [1, 2]]
-    self.check_schedule(xt, 5)
+    self.check_schedule(xt, 3)
     np.testing.assert_equal(xt.numpy(), (np.arange(16).reshape(4, 4))[[1, 2], [1, 2]])
 
   def test_advanced_indexing(self):
@@ -1753,13 +1754,13 @@ class TestIndexing(unittest.TestCase):
   def test_advanced_indexing_alt(self):
     X = Tensor.arange(6).reshape(3, 2)+1
     xt = X[[Tensor([2]), Tensor([1])]]
-    self.check_schedule(xt, 6)
+    self.check_schedule(xt, 3)
     np.testing.assert_equal(xt.numpy(), 6)
 
   def test_advanced_simple_indexing_combined(self):
     X = Tensor.arange(16).reshape(4, 4)
     xt = X[1:2, [1, 2]]
-    self.check_schedule(xt, 4)
+    self.check_schedule(xt, 2)
 
   def test_push_through_reshape(self):
     Tensor.manual_seed(0)
diff --git a/test/unit/test_llm_tokenizer.py b/test/unit/test_llm_tokenizer.py
index fca60bd7bb..7b65818a6f 100644
--- a/test/unit/test_llm_tokenizer.py
+++ b/test/unit/test_llm_tokenizer.py
@@ -1,7 +1,8 @@
-import unittest, base64, functools
+import unittest, base64, functools, sys
 from tinygrad.apps.llm import SimpleTokenizer, get_llama_re
 from tinygrad.helpers import fetch
 
+@unittest.skipIf(sys.platform == 'win32', "fetch race condition on Windows")
 class TestLLMTokenizer(unittest.TestCase):
   @functools.cached_property
   def basic_tok(self): return SimpleTokenizer(".*", { b"a": 0, b"b": 1, b"c": 2, b"ab": 3, b"bc": 4 }, { "<x>": 5, "<y>": 6, "<z>": 7 })
diff --git a/tinygrad/schedule/grouper.py b/tinygrad/schedule/grouper.py
index af62971b44..70682ad685 100644
--- a/tinygrad/schedule/grouper.py
+++ b/tinygrad/schedule/grouper.py
@@ -26,8 +26,8 @@ def realize_before_view(ctx:dict[UOp, None], view:UOp, tr:UOp) -> None:
 do_realize = PatternMatcher([
   # always realize SINK parents
   (UPat(Ops.SINK, name="s"), lambda ctx,s: ctx.update((x.base, None) for x in s.src if x.base.op not in ALWAYS_CONTIGUOUS)),
-  # always realize ASSIGN/CONTIGUOUS/GroupOp.Meta
-  (UPat({Ops.ASSIGN, Ops.CONTIGUOUS, *GroupOp.Meta}, name="tr"), realize),
+  # always realize ASSIGN/CONTIGUOUS/COPY/BUFFER_VIEW
+  (UPat({Ops.ASSIGN, Ops.CONTIGUOUS, Ops.COPY, Ops.BUFFER_VIEW}, name="tr"), realize),
   # realize before expand or unsafe pad ops
   (UPat(Ops.VIEW, src=(UPat(GroupOp.All-ALWAYS_CONTIGUOUS, name="tr"),), name="view"), realize_before_view),
   # realize parents of COPY, MSELECT, MSTACK