NVFP4 Emulation

[dsikka]

Summary:

• Add CompressedTensors NVFP4 Emulation Scheme
• Move emulations functionality into shared utilities

ModelOpt Emulation Changes:

• Don't run activations quantization to start --> seemed to be getting gibberish even with it turned off
• Update how the global scale is applied. I think they're storing the inverse? Updating this allows the coherent outputs for the Nvidia 70b checkpoint

Should now support ct models produced and compressed from the following branches:

1. FP4 Weights: [NVFP4][WIP] Add NVFp4 Support vllm-project/compressed-tensors#287
2. Compression: [WIP][NVFP4] Add compression/decompression code vllm-project/compressed-tensors#291
3. LLM Compressor: [NVFP4] Enable FP4 Weight-Only Quantization vllm-project/llm-compressor#1309

lm-evals/generations should work with weight only dequant:

lm_eval --model vllm \
    --model_args pretrained=nm-testing/llama2.c-stories110M-FP4,enforce_eager=True \
    --tasks gsm8k \
    --device cuda:0 \
    --batch_size 8

import numpy
import torch

from vllm import LLM, SamplingParams

prompts = ["The Swiss Alps are", "The president of the USA is", "The Boston Bruins are"]

# Create a sampling params object for greedy sampling
sampling_params = SamplingParams(temperature=0.80, top_p=0.95, max_tokens=40, min_tokens=10)
llm  = LLM('nm-testing/llama2.c-stories110M-FP4', enforce_eager=True)


# Print the outputs.
output = llm.generate(prompts, sampling_params)
for o in output:
    print(o.outputs[0].text)
    print("\n")

ToDo:

• Generally activation quant support - still need to understand how the input scales should be applied
• Improve compression speed in compressed-tensors. The current speed is a bottleneck atm

[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.

🚀