Support DeepSeekV3-style block FP8 quantization with CT

[mgoin]

Purpose

Redo of #20279

Relies on recent support in compressed-tensors (neuralmagic/compressed-tensors#372) and llm-compressor (vllm-project/llm-compressor#1607) to produce the models.

This PR implements W8A8 FP8 block quantization support for compressed-tensors models. This is focused on supporting the DeepSeekV3-style format, which has 128x128 block weights and 1x128 block activations (really per-token-group).

Most of the logic is ported directly from fp8.py and I hope to refactor the utilities to be shared eventually.

Test Plan

Manual testing with newly produced models. I'll add lm-eval in another PR

Test Result

Dense

CT result:

lm_eval --model vllm --model_args pretrained=mgoin/Qwen3-0.6B-FP8-BLOCK --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto
...
vllm (pretrained=mgoin/Qwen3-0.6B-FP8-BLOCK,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|↑  |0.3973|±  |0.0135|
|     |       |strict-match    |     5|exact_match|↑  |0.3995|±  |0.0135|

Ref:

lm_eval --model vllm --model_args pretrained=Qwen/Qwen3-0.6B-FP8 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto
...
vllm (pretrained=Qwen/Qwen3-0.6B-FP8,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|↑  |0.3973|±  |0.0135|
|     |       |strict-match    |     5|exact_match|↑  |0.3995|±  |0.0135|

MoE

CT result:

vllm (pretrained=mgoin/Qwen3-30B-A3B-FP8-BLOCK,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|↑  |0.8158|±  |0.0107|
|     |       |strict-match    |     5|exact_match|↑  |0.8923|±  |0.0085|

Ref:

vllm (pretrained=Qwen/Qwen3-30B-A3B-FP8,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|↑  |0.8158|±  |0.0107|
|     |       |strict-match    |     5|exact_match|↑  |0.8923|±  |0.0085|

[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.

🚀

[gemini-code-assist]

Code Review

This pull request adds support for DeepSeekV3-style block FP8 quantization using compressed-tensors. The changes are extensive, touching several files related to quantization and introducing new logic for handling block-quantized weights, especially in MoE layers. The PR also adds support for new hardware features like DeepGEMM on Blackwell.

While the overall approach seems correct, I've found a critical issue in the control flow of process_weights_after_loading in compressed_tensors_moe.py. The logic for block-quantized and non-block-quantized paths is mixed, leading to duplicated operations and potential runtime errors. This needs to be refactored to ensure correctness.

Other changes, such as refactoring to decouple layers and using more specific type hints, are good improvements to the codebase.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @mgoin.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork