[Bugfix] Fix FBGEMM integration

[mgoin]

FIX #16850 (comment)

Addresses issues with both the FP8 W8A8 and FP8 Marlin path.

Marlin needed to allow for the layer to not have weight_block_sizes present, and both paths needed to update their ignored modules check to work with fused modules.

Manually verified evals:

FP8 W8A8

WARNING 05-12 16:46:30 [marlin_utils_fp8.py:81] Your GPU does not have native support for FP8 computation but FP8 quantization is being used. Weight-only FP8 compression will be used leveraging the Marlin kernel. This may degrade performance for compute-heavy workloads.
vllm (pretrained=nm-testing/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform,enforce_eager=True,max_model_len=2048,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.7513|±  |0.0119|
|     |       |strict-match    |     5|exact_match|↑  |0.7536|±  |0.0119|

FP8 Marlin

vllm (pretrained=nm-testing/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform,enforce_eager=True,max_model_len=2048,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.7559|±  |0.0118|
|     |       |strict-match    |     5|exact_match|↑  |0.7574|±  |0.0118|

====

Original failure found in CI: https://buildkite.com/vllm/ci/builds/19754#0196bd80-d18b-4c8a-bf76-50e9a53c5a6c/43-353

ERROR 05-10 22:15:39 [multiproc_executor.py:487]   File "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/layers/quantization/fbgemm_fp8.py", line 144, in process_weights_after_loading
ERROR 05-10 22:15:39 [multiproc_executor.py:487]     prepare_fp8_layer_for_marlin(layer)
ERROR 05-10 22:15:39 [multiproc_executor.py:487]   File "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/layers/quantization/utils/marlin_utils_fp8.py", line 122, in prepare_fp8_layer_for_marlin
ERROR 05-10 22:15:39 [multiproc_executor.py:487]     if layer.weight_block_size is None:
ERROR 05-10 22:15:39 [multiproc_executor.py:487]        ^^^^^^^^^^^^^^^^^^^^^^^
ERROR 05-10 22:15:39 [multiproc_executor.py:487]   File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1940, in __getattr__
ERROR 05-10 22:15:39 [multiproc_executor.py:487]     raise AttributeError(
ERROR 05-10 22:15:39 [multiproc_executor.py:487] AttributeError: 'QKVParallelLinear' object has no attribute 'weight_block_size'

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[jinzhen-lin]

The second CI failure seems a different problem of compressed-tensors.

[robertgshaw2-redhat]

Thanks Michael

[mgoin]

@jinzhen-lin This actually is not sufficient. The FBGEMM model runs but has no correctness

lm_eval --model vllm --model_args pretrained=nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform,enforce_eager=True,max_model_len=2048,tensor_parallel_size=4 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto

vllm (pretrained=nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform,enforce_eager=True,max_model_len=2048,tensor_parallel_size=4,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |    0|±  |     0|
|     |       |strict-match    |     5|exact_match|↑  |    0|±  |     0|

[jinzhen-lin]

@jinzhen-lin This actually is not sufficient. The FBGEMM model runs but has no correctness

lm_eval --model vllm --model_args pretrained=nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform,enforce_eager=True,max_model_len=2048,tensor_parallel_size=4 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto

vllm (pretrained=nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform,enforce_eager=True,max_model_len=2048,tensor_parallel_size=4,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |    0|±  |     0|
|     |       |strict-match    |     5|exact_match|↑  |    0|±  |     0|

I would find the reason tomorrow.

[mgoin]

I actually found I get the same result using FBGEMM models with or without marlin (i.e. W8A8). There is something broken in the base case

EDIT: This was due to incorrect skip module checks