[thunderstornX]

Overview

Fixes #58742. Resolved quantization names missing from the estimator lookup silently drop attention and FFN work from MFU/MBU estimates.

Claims

• Resolve weight size from GPTQ/AWQ bit widths, online quantization specifications, and mixed linear/expert precision.
• Retain components with a deduplicated warning and model-dtype estimate when an unsupported method or checkpoint layer selection cannot be represented by one width.
• Preserve existing supported methods and unquantized controls.

Validation

Rebased onto current main 276fbcf, resolving the upstream move of Self into typing. Implementation and test bodies are unchanged from the previously tested branch.

OMP_NUM_THREADS=1 .venv/bin/python -m pytest --confcutdir=tests/v1/metrics tests/v1/metrics/test_perf_metrics.py -q
.venv/bin/pre-commit run --files vllm/v1/metrics/perf.py tests/v1/metrics/test_perf_metrics.py

100 metrics tests passed. All applicable pre-commit checks passed, including the repository's minimum-version mypy check.

Executed 16 real checkpoint/flag combinations six times against the rebased stack: 96 online runs passed, each evaluating prefill and decode. A separate 16-case run used the unpatched current-main estimator with identical checkpoint revisions, configuration hashes, and resolved layer lists. All repeats were identical; unaffected controls and the previously approved estimates were unchanged.

Checkpoint	Upstream decode GFLOPs	Patched decode GFLOPs
GPT-OSS-20B	1.158	7.436
Qwen2.5 GPTQ/AWQ	0.272	1.076
Qwen3 bf16 control	1.427	1.427

Validation used Python 3.12 and CPU Torch with actual ModelConfig, VllmConfig._get_quantization_config, and estimator execution. No weights were loaded or GPU inference run; model outputs, kernels, and serving behavior are unchanged.

Details

Both attention and FFN parsers share weight-size resolution. GPT-OSS retains bf16 linear layers and FP4 expert accounting; DeepSeek-V4's FFN accounting is restored while its separate unsupported attention architecture remains outside this fix.

The fallback can overestimate weight traffic and does not provide per-layer accounting. Automatically discovered layer lists in the tested Qwen GPTQ/AWQ checkpoints trigger the documented 2-byte fallback rather than assuming every projection is 4-bit.

No overlapping fix was found in issue-number or area searches. Related #58745, #58836, and #54228 address speculative-token accounting, hybrid attention, and expert activation counts respectively.

Pull Request Checklist

• Design Fit: confines changes to the analytic estimator and shares resolution between its parsers.
• Testing and Validation: regression coverage and real configuration execution passed; upstream CI remains maintainer-authorized.
• Code Quality and Style: current pre-commit and type checks passed.
• Pull Request Contents: issue, reproducer commands, results, and limitations are included.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[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.

PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment /ci run for upstream CI or /amd-ci run for AMD CI only whenever CI signals are needed.

Once the PR is approved or has the ready label, the PR author can also use the corresponding /ci run, /ci retry, and /ci cancel commands, or their /amd-ci variants. New commits do not start upstream CI automatically.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[markmc]

@thillai-c ptal