[Bugfix] Remove NVFP4 scales assertions to fix load_format=dummy

[mgoin]

Currently on 0.9.0 when running vllm serve nvidia/DeepSeek-R1-FP4 --quantization modelopt_fp4 -tp 8 --load-format dummy on A100, we run into

(VllmWorker rank=5 pid=3888806) ERROR 05-28 18:39:46 [multiproc_executor.py:487]   File "/home/mgoin/venvs/vllm-rel/lib/python3.12/site-packages/vllm/model_executor/layers/quantization/utils/marlin_utils_fp4.py", line 25, in fp4_marlin_process_scales
(VllmWorker rank=5 pid=3888806) ERROR 05-28 18:39:46 [multiproc_executor.py:487]     assert (marlin_scales >= 0).all()
(VllmWorker rank=5 pid=3888806) ERROR 05-28 18:39:46 [multiproc_executor.py:487]            ^^^^^^^^^^^^^^^^^^^^^^^^^^
(VllmWorker rank=5 pid=3888806) ERROR 05-28 18:39:46 [multiproc_executor.py:487] AssertionError

Then on B200 when running vllm serve nvidia/DeepSeek-R1-FP4 --quantization modelopt_fp4 -tp 4 --load-format dummy, we run into

(VllmWorker rank=3 pid=2693151) ERROR 05-28 18:48:28 [multiproc_executor.py:487]   File "/home/mgoin/code/vllm/vllm/model_executor/layers/quantization/modelopt.py", line 588, in process_weights_after_loading
(VllmWorker rank=3 pid=2693151) ERROR 05-28 18:48:28 [multiproc_executor.py:487]     assert torch.allclose(
(VllmWorker rank=3 pid=2693151) ERROR 05-28 18:48:28 [multiproc_executor.py:487]            ^^^^^^^^^^^^^^^
(VllmWorker rank=3 pid=2693151) ERROR 05-28 18:48:28 [multiproc_executor.py:487] AssertionError: w1_weight_scale_2 must match w3_weight_scale_2

These asserts are hit because dummy weights are likely to have random values. It doesn't seem this check is truly necessary

[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.

🚀

[mgoin]

PTAL @pavanimajety @jinzhen-lin

[jinzhen-lin]

In Marlin NVFP4, I changed the scales from FP8-S1E4M3 to a special FP8-S0E5M3 format to speedup the dequantization, this assumed the origin scales to be >=0, if we can always ensure that for real weight, this assertion can be removed.