Bitcast p2 bfloat16 tests + clang fix (#2635)

* add bf16 test support

this model takes me almost a minute to download though:

https://huggingface.co/TinyPixel/Llama-2-7B-bf16-sharded/resolve/main/pytorch_model-00001-of-00014.bin?download=true: 100%|█████████████████████████████| 981M/981M [00:40<00:00, 24.2MB/s]

* ensure we first load if it is bitcast to avoid taking the address of an rvalue

* tiny bf16 in the cloud

skip GPU

* should skip torch

lint

* Revert "ensure we first load if it is bitcast to avoid taking the address of an rvalue"

This reverts commit b86a28ab84bc1173764b2d480218e8de41a32390.

* break the kernel

* skip LLVM and GPU in CI

* skip CUDA

test/unit/test_disk_tensor.py

@@ -3,16 +3,17 @@ import unittest
 import numpy as np
 from tinygrad.tensor import Tensor, Device
 from tinygrad.nn.state import safe_load, safe_save, get_state_dict, torch_load
-from tinygrad.helpers import fetch, temp
+from tinygrad.helpers import CI, fetch, temp
 from tinygrad.helpers import Timing
 
 def compare_weights_both(url):
   import torch
   fn = fetch(url)
   tg_weights = get_state_dict(torch_load(fn))
-  torch_weights = get_state_dict(torch.load(fn), tensor_type=torch.Tensor)
+  torch_weights = get_state_dict(torch.load(fn, map_location=torch.device('cpu')), tensor_type=torch.Tensor)
   assert list(tg_weights.keys()) == list(torch_weights.keys())
   for k in tg_weights:
+    if torch_weights[k].dtype == torch.bfloat16: torch_weights[k] = torch_weights[k].float() # numpy doesn't support bfloat16
     np.testing.assert_equal(tg_weights[k].numpy(), torch_weights[k].numpy(), err_msg=f"mismatch at {k}, {tg_weights[k].shape}")
   print(f"compared {len(tg_weights)} weights")
 
@@ -23,6 +24,14 @@ class TestTorchLoad(unittest.TestCase):
   def test_load_enet_alt(self): compare_weights_both("https://download.pytorch.org/models/efficientnet_b0_rwightman-3dd342df.pth")
   # pytorch zip format
   def test_load_convnext(self): compare_weights_both('https://dl.fbaipublicfiles.com/convnext/convnext_tiny_1k_224_ema.pth')
+
+  # for GPU, cl_khr_fp16 isn't supported
+  # for LLVM, it segfaults because it can't link to the casting function
+  # CUDACPU architecture is sm_35 but we need at least sm_70 to run fp16 ALUs
+  @unittest.skipIf(Device.DEFAULT in ["GPU", "LLVM", "CUDA"] and CI, "fp16 broken in some backends")
+  @unittest.skipIf(Device.DEFAULT == "TORCH", "torch doesn't support the way we load bfloat (cast to uint32)")
+  def test_load_llama2bfloat(self): compare_weights_both("https://huggingface.co/qazalin/bf16-lightweight/resolve/main/consolidated.00.pth?download=true")
+
   # TODO: support pytorch tar format with minimal lines
   #def test_load_resnet(self): compare_weights_both('https://download.pytorch.org/models/resnet50-19c8e357.pth')
 

tinygrad/nn/state.py

@@ -77,7 +77,7 @@ def torch_load(fn:str):
     # upstream LLaMA also does this conversion:
     # https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/generation.py#L95
     # TODO: should this be done in the example instead? or maybe we don't need this anymore with better bfloat16 support
-    if storage[1] == dtypes.bfloat16: ret = ret.cast(dtypes.uint16).to(Device.DEFAULT).cast(dtypes.uint32).mul(1<<16).bitcast(dtypes.float32).half()
+    if storage[1] == dtypes.bfloat16: ret = ret.cast(dtypes.uint16).to(Device.DEFAULT).cast(dtypes.uint32).mul(1<<16).contiguous().bitcast(dtypes.float32).half()
     else: ret = ret.cast(storage[1])
 
     # 7 lines to deal with permuted tensors. NOTE: this currently requires reading off the disk