[Sync] dev/perf sync with upstream 20260130

[kliuae-amd]

Purpose

Sync changes from upstream.

Upstream vLLM cf1167e (v0.15.0 release)
aiter: 6af8b6874

Test Plan

Evaluate the following models of interest.
LLM: gsm8k
VLM: ChartQA

Test Result

LLM

deepseek-ai/DeepSeek-R1, TP8

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.9553|_  |0.0057|
|     |       |strict-match    |     5|exact_match|_  |0.9553|_  |0.0057|

deepseek-ai/DeepSeek-R1, TP8 + shared expert fusion

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.9553|_  |0.0057|
|     |       |strict-match    |     5|exact_match|_  |0.9545|_  |0.0057|

deepseek-ai/DeepSeek-R1, TP8 + EP

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.9553|_  |0.0057|
|     |       |strict-match    |     5|exact_match|_  |0.9545|_  |0.0057|

deepseek-ai/DeepSeek-R1, TP8 + EP + shared expert fusion

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.9507|_  | 0.006|
|     |       |strict-match    |     5|exact_match|_  |0.9492|_  | 0.006|

EmbeddedLLM/deepseek-r1-FP8-Dynamic PTPC FP8, TP8

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.9378|_  |0.0067|
|     |       |strict-match    |     5|exact_match|_  |0.9378|_  |0.0067|

EmbeddedLLM/deepseek-r1-FP8-Dynamic PTPC FP8, TP8 + EP

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.9553|_  |0.0057|
|     |       |strict-match    |     5|exact_match|_  |0.9560|_  |0.0056|

EmbeddedLLM/Qwen3-Coder-480B-A35B-Instruct-FP8-Dynamic, TP4

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.8779|_  |0.0090|
|     |       |strict-match    |     5|exact_match|_  |0.8491|_  |0.0099|

EmbeddedLLM/Qwen3-Coder-480B-A35B-Instruct-FP8-Dynamic, TP4 + EP

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.8840|_  |0.0088|
|     |       |strict-match    |     5|exact_match|_  |0.8529|_  |0.0098|

Qwen/Qwen3-Next-80B-A3B, TP4

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.8552|_  |0.0097|
|     |       |strict-match    |     5|exact_match|_  |0.8089|_  |0.0108|

Qwen/Qwen3-Omni-30B-A3B-Instruct, TP2

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|_  |0.8597|_  |0.0096|
|     |       |strict-match    |     5|exact_match|_  |0.8476|_  |0.0099|

VLM

Qwen/Qwen3-VL-235B-A22B-Instruct TP4

{
    "explicit_prompt_relaxed_correctness": 0.8652,
    "anywhere_in_answer_relaxed_correctness": 0.8672
}

RedHatAI/Qwen3-VL-235B-A22B-Instruct-FP8-dynamic TP4

{
    "explicit_prompt_relaxed_correctness": 0.868,
    "anywhere_in_answer_relaxed_correctness": 0.8688
}

Qwen/Qwen2.5-VL-72B-Instruct

{
    "explicit_prompt_relaxed_correctness": 0.8648,
    "anywhere_in_answer_relaxed_correctness": 0.8864
}

RedHatAI/Qwen2.5-VL-72B-Instruct-FP8-dynamic TP2

{
    "explicit_prompt_relaxed_correctness": 0.8744,
    "anywhere_in_answer_relaxed_correctness": 0.8888
}

Qwen/Qwen3-Omni-30B-A3B-Instruct TP2

{
    "explicit_prompt_relaxed_correctness": 0.8712,
    "anywhere_in_answer_relaxed_correctness": 0.872
}

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.