[eble-amd]

Purpose

Merge b131311 from upstream to gfx11 in stages. Conflict resolution is recorded in my merge commits.

Test Plan & Result (agentic)

Run date: 2026-10-06, Strix Halo (gfx1151). ROCm/torch/Triton rocm10.2.0a20261006 on both sides.

• A (baseline): 34a622ef5b1b (gfx11 after PR Merge KV cache refactor from upstream #1344)
• B (candidate): 313fad5a83ee, merges of upstream 7156c63bef, 31739ceb0b, f79a2f5582,
  7c877062ac, b131311fb4
• Same board, back-to-back, one run per case per side.

Verdict

5/5 PASS on both sides. All TTFT and decode deltas are within ±2.2%, i.e. run-to-run noise.
Input/output token counts and spec-decode acceptance are identical except one row (see notes).
Nothing in this set flags the merge.

A/B comparison (B relative to A)

case	status	A TTFT ms	B TTFT ms	Δ%	A dec tok/s	B dec tok/s	Δ%	accept. len A / B	runner A → B
Qwen3.6-35B-A3B-W4A16-llmcompressor_VLM_4ksys_1k_320x240_100out_dyn-lm-int8g32	PASS	508.5	508.2	-0.1	80.44	80.33	-0.1	–	V1 → V2
Qwen3.6-35B-A3B-W4A16-llmcompressor-mtpfix_VLM_dyn-lm-int8g32_MTP-mmstar	PASS	221.0	216.2	-2.2	99.33	99.33	0.0	1.85 / 1.85	V1 → V2
Qwen3-30B-A3B-Instruct-2507-AWQ-4bit_specdec-eagle3-aime	PASS	221.8	221.9	0.0	70.52	71.05	+0.8	2.59 / 2.59	V1 → V2
Gemma-4-E4B-IT_VLM_AWQ-INT4_DFlare-B16-K15_RoboVQA	PASS	303.6	305.0	+0.5	166.32	166.14	-0.1	6.91 / 6.91	V2 → V2
Gemma-4-31B-IT_VLM_OpenNav_AWQ-4bit_MTP-mmstar	PASS	1220.0	1220.0	0.0	27.57	27.49	-0.3	2.90 / 2.90	V2 → V2

Built-in sanity checks (1+1 / image colour) passed on both sides for every row except DFlare,
which skips them by config; DFlare instead requires spec-decode acceptance length ≥ 5.0, which
passed.

Notes

• Cases: chosen for the merged changes most likely to reach gfx1151: upstream [Model Runner V2] Use MRV2 for all models by default vllm-project/vllm#53183
  makes Model Runner V2 the default for MoE/hybrid models (three rows switch V1 → V2), plus
  spec-decode changes (MTP, eagle3, DFlare) and multimodal processing refactors.
• Qwen3.6 MTP-mmstar output tokens 1280 (A) vs 1267 (B). A build of the same source as A
  produced 1267 in an earlier run, so this is run-to-run variation, not the merge.

---

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.