[ Misc ] Expand Fp8 MoE Support to Qwen #6088
Conversation
|
@robertgshaw2-neuralmagic we are also suffering from the illegal memory access even before this refactoring. It's weird because I didn't find this issue at v0.5.0 and it's still working for me on certain H100 instances. There are some possible clues but I haven't verified:
I wish I could allocate some bandwidth for this but no guaranteed. We could sync offline for updates. |
@comaniac Okay that would be great. |
|
@robertgshaw2-neuralmagic @comaniac There is a potential risk of illegal memory access, I have made changes but have not yet submitted them. Please refer to:add_device_gurad |
Interesting. Do you think the potential illegal memory access is due to the wrong device being used? |
I'm not entirely sure if we're talking about the same thing.I was testing whether the triton kernel could run on non-zero GPUs and found this error. Below, I provide my check steps, which may be helpful. 1. Create python snippet
import pytest
import torch
from transformers import MixtralConfig
from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
from vllm.model_executor.models.mixtral import MixtralMoE
CUDA_DEVICES = [
f"cuda:{1}" ]
@torch.inference_mode()
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_mixtral_moe(device,dtype: torch.dtype=torch.float16):
"""Make sure our Mixtral MoE implementation agrees with the one from
huggingface."""
# Instantiate our and huggingface's MoE blocks
config = MixtralConfig()
hf_moe = MixtralSparseMoeBlock(config).to(dtype).to(device)
vllm_moe = MixtralMoE(
num_experts=config.num_local_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.intermediate_size,
params_dtype=dtype,
tp_size=1,
).to(device)
# Load the weights
vllm_moe.gate.weight.data[:] = hf_moe.gate.weight.data
for i in range(config.num_local_experts):
weights = (hf_moe.experts[i].w1.weight.data,
hf_moe.experts[i].w3.weight.data)
vllm_moe.experts.w13_weight[i][:] = torch.cat(weights, dim=0)
vllm_moe.experts.w2_weight[i][:] = hf_moe.experts[i].w2.weight.data
# Generate input batch of dimensions [batch_size, seq_len, hidden_dim]
hf_inputs = torch.randn((1, 64, config.hidden_size)).to(dtype).to(device)
# vLLM uses 1D query [num_tokens, hidden_dim]
vllm_inputs = hf_inputs.flatten(0, 1)
# Run forward passes for both MoE blocks
hf_states, _ = hf_moe.forward(hf_inputs)
vllm_states = vllm_moe.forward(vllm_inputs)
mixtral_moe_tol = {
torch.float32: 1e-3,
torch.float16: 1e-3,
torch.bfloat16: 1e-2,
}
assert torch.allclose(hf_states.flatten(0, 1),
vllm_states,
rtol=mixtral_moe_tol[dtype],
atol=mixtral_moe_tol[dtype])2 Run testing pytest test_moe.pyIn my device(A800), will raise the following error: 3. Check this error Then I utilize compute-sanitizer --tool memcheck --target-processes all pytest test_moe.pyand the key log : 4 Fix PS : My initial intention was still testing triton, so I didn't submit this changes. If it proves useful, I can submit it asap. |
|
Thanks for the detail steps, which are helpful. In the e2e case I believe vllm would make sure all tensors are on the right device, so this shouldn't be an issue, but I'll keep in mind. |
robertgshaw2-neuralmagic
requested review from
comaniac,
simon-mo,
mgoin,
pcmoritz and
WoosukKwon
Co-authored-by: Michael Goin <[email protected]>
Co-authored-by: Michael Goin <[email protected]>
|
/ready |
github-actions
bot
added
the
ready
|
Have to remove l4 tests. Triton kernel requires H100 |
|
有希望在下一版本支持Qwen2 MOE 量化版本吗? |
Pull request was closed
|
Closed because this functionality was completed in fb6af8b |
SUMMARY:
BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE
PR Checklist (Click to Expand)
Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.
PR Title and Classification
Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:
[Bugfix]for bug fixes.[CI/Build]for build or continuous integration improvements.[Doc]for documentation fixes and improvements.[Model]for adding a new model or improving an existing model. Model name should appear in the title.[Frontend]For changes on the vLLM frontend (e.g., OpenAI API server,LLMclass, etc.)[Kernel]for changes affecting CUDA kernels or other compute kernels.[Core]for changes in the core vLLM logic (e.g.,LLMEngine,AsyncLLMEngine,Scheduler, etc.)[Hardware][Vendor]for hardware-specific changes. Vendor name should appear in the prefix (e.g.,[Hardware][AMD]).[Misc]for PRs that do not fit the above categories. Please use this sparingly.Note: If the PR spans more than one category, please include all relevant prefixes.
Code Quality
The PR need to meet the following code quality standards:
format.shto format your code.docs/source/if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.Notes for Large Changes
Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with
rfc-requiredand might not go through the PR.What to Expect for the Reviews
The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:
action-requiredlabel on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.Thank You
Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!