Dtype mismatch in linear layer

[tfogal]

🐛 Bug

[rank0]:   File "/home/tfogal/dev/thunder/thunder/core/langctxs.py", line 132, in _fn
[rank0]:     result = fn(*args, **kwargs)
[rank0]:   File "/home/tfogal/dev/thunder/thunder/torch/__init__.py", line 4580, in linear
[rank0]:     return prims.linear(a, w, bias)
[rank0]:   File "/home/tfogal/dev/thunder/thunder/core/symbol.py", line 264, in __call__
[rank0]:     result = self.meta(*args, **kwargs)
[rank0]:   File "/home/tfogal/dev/thunder/thunder/core/langctxs.py", line 132, in _fn
[rank0]:     result = fn(*args, **kwargs)
[rank0]:   File "/home/tfogal/dev/thunder/thunder/core/prims.py", line 3585, in linear_meta
[rank0]:     utils.check(
[rank0]:   File "/home/tfogal/dev/thunder/thunder/core/baseutils.py", line 103, in check
[rank0]:     raise exception_type(s())
[rank0]: RuntimeError: Expected a.dtype=bfloat16 and w.dtype=float32 to be the same!

Full log of the run that includes the unabbreviated traceback.

To Reproduce

HYDRA_FULL_ERROR=1 \
THUNDER_ANNOTATE_TRACES=1 \
NEMO_THUNDER_NEVA=1 \
python3 ./examples/multimodal/multimodal_llm/neva/neva_pretrain.py trainer.precision=16 model.megatron_amp_O2=False trainer.num_nodes=1 trainer.devices=1 trainer.val_check_interval=10 trainer.limit_val_batches=5 trainer.log_every_n_steps=1 ++exp_manager.max_time_per_run=00:00:03:00 trainer.max_steps=20 model.micro_batch_size=2 model.global_batch_size=4 model.tensor_model_parallel_size=1 model.pipeline_model_parallel_size=1 exp_manager.create_checkpoint_callback=False model.data.data_path=./data/multimodal/tiny-neva/dummy.json model.data.image_folder=./data/multimodal/tiny-neva/images model.tokenizer.library=sentencepiece model.tokenizer.model=./data/multimodal/tiny-neva/tokenizer_add_special.model model.num_layers=2 model.hidden_size=5120 model.ffn_hidden_size=13824 model.num_attention_heads=40 model.normalization=rmsnorm model.data.num_workers=0 model.data.conv_template=llama_2 model.mm_cfg.vision_encoder.from_pretrained=openai/clip-vit-large-patch14 model.mm_cfg.llm.from_pretrained=null model.use_flash_attention=false exp_manager.exp_dir=./foo-neva-train

Note you'll need the referenced ./data directory.

Expected behavior

Environment

$ nvidia-smi | grep -i cuda
| NVIDIA-SMI 555.42.02              Driver Version: 555.42.02      CUDA Version: 12.5     |
$ python3 -m pip freeze | egrep -i "(nvfuser)|(lightning)|(thunder)|(nemo)|(megatron)|(torch)"
-e git+ssh://git@github.com/tfogal/lightning.git@8df5db52ead1804f9021bb07caa2d4a7a6ab03a1#egg=lightning
lightning-cloud==0.5.69
-e git+ssh://git@github.com/Lightning-AI/lightning-thunder.git@72e033a0e0dfe44d4770dec2399a9058971003ec#egg=lightning_thunder
lightning-utilities==0.11.2
megatron_core @ file:///home/tfogal/Megatron-LM
-e git+ssh://git@github.com/NVIDIA/NeMo.git@c86449e1a93049d2283ebcea8ee4546f2ea241de#egg=nemo_toolkit
# Editable Git install with no remote (nvfuser==0.2.6+git9c5f006)
-e /opt/pytorch/nvfuser
open-clip-torch==2.24.0
pytorch-lightning==2.3.0
-e git+https://github.com/pytorch/pytorch.git@bd72e28314d8d63bb347becb8309f5ac7761c6b5#egg=torch
torchdiffeq==0.2.4
torchmetrics==1.4.0.post0
torchsde==0.2.6
torchvision @ git+https://github.com/pytorch/vision.git@bf01bab6125c5f1152e4f336b470399e52a8559d
-e git+https://gitlab-ci-token:glcbt-64_VRyDQgDXFf-uV3J9S3gy@gitlab-master.nvidia.com/dl/pytorch/update-scripts.git@5bbcbd6d7aff52c6e6d0b47febe053d4894b3464#egg=zpyt_nightly_ci

cc @crcrpar @tfogal

[t-vi]

[rank0]:   File "/home/tfogal/dev/nemo/nemo/collections/nlp/modules/common/megatron/adapters/parallel_adapters.py", line 914, in forward
[rank0]:     return self.mm_projector(x)

we could likely print self.mm_projector.weight.dtype and x.dtype to figure out what we get without dtype.

[kshitij12345]

I am seeing a different error related to advanced indexing.

[rank0]:   File "thunder/core/proxies.py", line 1333, in __getitem__
[rank0]:     return method(self, key)
[rank0]:   File "thunder/core/symbol.py", line 268, in __call__
[rank0]:     result = self.meta(*args, **kwargs)
[rank0]:   File "thunder/core/langctxs.py", line 132, in _fn
[rank0]:     result = fn(*args, **kwargs)
[rank0]:   File "thunder/torch/__init__.py", line 890, in getitem
[rank0]:     return clang.getitem(a, key)
[rank0]:   File "thunder/core/langctxs.py", line 132, in _fn
[rank0]:     result = fn(*args, **kwargs)
[rank0]:   File "thunder/clang/__init__.py", line 868, in getitem
[rank0]:     return _advanced_indexing(a, key)
[rank0]:   File "thunder/core/langctxs.py", line 132, in _fn
[rank0]:     result = fn(*args, **kwargs)
[rank0]:   File "thunder/clang/__init__.py", line 729, in _advanced_indexing
[rank0]:     utils.check(
[rank0]:   File "thunder/core/baseutils.py", line 103, in check
[rank0]:     raise exception_type(s())
[rank0]: RuntimeError: Advanced indexing currently only supports zero or one-dimensional integer tensors, but found a tensor with dtype int64 and 2 dimensions

thunder commit used - 72e033a

Full Log: neva.log

[tfogal]

triage review

• @tfogal dig in and see if latest thunder (or 72e033a) reproduces the indexing error or something else
• we may have a type promotion that PyTorch does not have. let's print some dtypes at various levels to see what's going on here
• we might also be not mimicking some autocast logic
• This seems to happen to us a bunch through recurrent paths

[kshitij12345]

I have been able to repro the failure with an independent script. The failure happens due to the interaction of autocast and mixed input dtypes.

import thunder
import torch

def foo(x, w):
    return torch.nn.functional.linear(x, w)

device = torch.device("cuda")
with device:
    # Mixed input types.
    x, w = torch.randn(16, 16, dtype=torch.bfloat16), torch.randn(16, 16)

    # Same input types (works with thunder)
    # x, w = torch.randn(16, 16), torch.randn(16, 16)

    print(x.dtype, w.dtype)
    
with torch.autocast("cuda", torch.bfloat16):
    # Eager autocast handles mixed input types.
    eager_out = foo(x, w)

    # `thunder.jit` doesn't handle mixed inputs.    
    jfoo = thunder.jit(foo)
    jit_out = jfoo(x, w)


print(thunder.last_traces(jfoo)[-1])
torch.testing.assert_close(eager_out, jit_out)

[tfogal]

I have been able to repro the failure with an independent script.

Great! Thank you, excellent work :-)

[kshitij12345]

The reason it fails currently is because, while tracing with thunder.jit -

1. We first try to generate the computation trace based on the actual function/model.
2. And we apply the autocast transform later on the above computation trace.

With mixed input dtypes, we fail at step 1 as these operators don't allow mixed inputs.

(In eager, with the context manager active, dispatcher first applies the conversion before passing the converted inputs to the operators).

Potential Fix:

1. We should stash the detail that autocast is enabled in CompileData.
2. Then while tracing at the place where we translate torch functions to thunder - we should check if autocast is active and if we have a rule for this function.

@t-vi I would like your opinion on the same and some pointers. Thank you!

[t-vi]

Great analysis @kshitij12345 !

For 1: We do have autocast handling in thunder.jit and cache_info.

lightning-thunder/thunder/__init__.py

Line 395 in da23a0b

cache_info["is_autocast_enabled"] = is_autocast_enabled

For 2: To my mind, this is a thunder.torch thing more than something specific to jit_ext, so I would probably look at trying to handle it in thunder.torch.torchsymbol

lightning-thunder/thunder/torch/__init__.py

Line 108 in da23a0b

def __call__(self, fn: Callable) -> Symbol:

WDYT?