perf(GFX1100-TG200): T2b flips ROCm support_static_graph_mode

[ghazni101]

Closes #2774.

Row: GFX1100-TG200

The last false predicate in the dense decode-graph gate chain falls:
support_static_graph_mode() returns true on ROCm. With the hipGraph
capture/replay seam already implemented (rocm_backend.hip, with the
mutate-src-then-replay assertion) and the keep-quant scratch pool already
capture-safe (hipMallocAsync, stream-ordered, never freed), the flip engages
Qwen3_5DenseDecodeGraph for uniform decode steps.

Verified live: graph capture and replay engaged on a 16-token run; output
coherent and token-identical to the upstream baseline (Qwen3.5-4B Q4_K,
32-token greedy decode, seed 0).

A/B (acceptance workload, 256 tok x 5): 36.4 tok/s median in a co-tenancy
window vs 35.8 same-window split-arm baseline -- neutral-to-slightly-positive
under contention, as expected since a shared GPU queue partially hides the
dispatch gap.

Evidence: docs/bench-evidence/gfx1100-tg200-t2b-20260823.md

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:GLM-5-2 [OMP]

[VikashLoomba]

Current-main audit at c796fea41f74fe90b8cf78190eeb2a5b4c977449 confirms the activation gap, but the cumulative evidence is insufficient for a current merge.

RocmBackend::SupportsGraphCapture() is already true at rocm_backend.hip:337. RocmPlatform still inherits false at rocm.cpp:91-97. Active vLLM e126687a9 returns true at vllm/platforms/rocm.py:1088-1090. The platform override is independently derivable; the existing backend, scratch retirement, pool pinning, pre-grow, and (S,q,spec) keys are already on main.

The override activates all admitted consumers, not only the historical 4B single-stream run. Current evidence needs a real ROCm test entering through ModelRegistry::Forward, with capture/replay counters and changing input tokens and metadata. Directly constructed CPU graph-driver tests fake replay. The backend tests cover mutable-source D2D copies and one warmed GEMM. They do not prove default pageable H2D metadata refresh, retained scratch across shapes, or recapture and teardown with outstanding work.

Reconcile ownership before the new implementation claim: issue #2774 and this PR name GFX1100-TG200, while the current campaign spec names BACKEND-ROCM and points to closed records-only #2427. Keep #2774 as the existing activation issue, with BACKEND-ROCM ownership and T2b campaign attribution. Issue #332's absent-backend-seam description also needs updating; its remaining activation gap stays open. Existing attention wiring debt #2923 remains separate.

Commit a bounded current-main spec before code, with active upstream tests and production-callsite mutations. Preserve historical negative and contended evidence rather than removing it. Acceptance still needs the #2856 correctness prerequisite, runnable active-pin ROCm oracle, and locked hardware gates. No current safety or speed result is claimed by this audit.

[localai-org-maint-bot]

Held back rather than bounced — nothing here is wrong, and I had it in the merge
set until the last check. Two things changed my mind, and both are about blast
radius rather than about the flip being a bad idea.

The flip is live by default on every ROCm arch, not just gfx1100.
support_static_graph_mode() is per-platform, and the drivers conjunct it with
vt::GraphCaptureEnabled() — which reads VLLM_CPP_CUDAGRAPH and defaults
on (e == nullptr || e != "0") — and SupportsGraphCapture(). So with this
merged, a Qwen3, Qwen3.5, Qwen3-MoE, DeepSeek-V2, Voxtral or DFlash decode arms
the graph on gfx1151, gfx1103, gfx1200 and gfx1201 too, with no per-arch guard
and no evidence on any of them. docs/bench-evidence/gfx1100-tg200-t2b-20260823.md
is a gfx1100 measurement.

I could not gate the part that matters. strix:gpu0 (gfx1151) is the only
AMD board on the fleet and I ran the ROCm suite there, but
test_rocm_backend.cpp exercises the capture seam — Begin/End/Replay — and
nothing in the suite drives a model decode graph. So a green run on strix would
have proved the primitives work and said nothing about the predicate this
changes. I would rather say that than merge on a caveat that reads like
coverage.

The third reason is circumstantial but real: the campaign this flip belongs to
is bounced on #2790
for four undisclosed removals. Landing a global default-on change whose
justification is that campaign's decode work, while the decode work itself is
not landing, takes on the risk without anyone getting the benefit today.

What would land it, in rough order of how much work each is:

1. Scope the predicate the way feat(BACKEND-TENSTORRENT-HOST-FREE-FORWARD): make captured decode the TT default for Qwen3-dense #2910 just scoped the Tenstorrent one — that
   pull request added Platform::static_graph_requires_opt_in() returning false
   in the base, so a platform can default capture on for exactly the
   architectures that carry gate evidence. The same shape here would let
   gfx1100 default on and leave the other four arches opt-in.
2. Or a decode-graph gate that runs on a non-gfx1100 ROCm board, so the flip has
   coverage somewhere the fleet can actually reach.
3. Or land it with perf(BACKEND-ROCM): add opt-in MMVQ decode GEMV #2790's stack once that is repaired, where it is measured as
   part of the thing it was measured with.

One separate item, worth fixing whichever way this goes:
#2943. The evidence-file edit
drops the whole "Session-state note (hindsight unavailable)" paragraph. The
stale preamble deserved to go; the last sentence in it did not — the
pointer-keyed activation-quant cache being unsound under the DevicePool
block-recycling allocator is a closed negative, and that paragraph is the only
place the tree records it. Keep the finding, drop the preamble.

#2778 and #2782 from your set are landing today.

[ghazni101]

Reworked per your conditions. Two commits on b9f2ef4 (the external-contributor landing branch): the bounded spec FIRST (ba706df, .agents/specs/gfx1100-tg200-t2b-static-graph-opt-in.md), then the code (62f7739): the flip is scoped per-arch via Platform::static_graph_requires_opt_in() mirroring #2910's Tenstorrent pattern — gfx1100 admits default capture; gfx1151/1103/1200/1201 keep eager until they carry their own evidence (rocm.cpp:106-130, IsGfx1100 helper at 204-211). The #2943 pointer-keyed activation-quant-cache sentence is preserved in the evidence file (the rework resets to the base that has it). New self-skipping test test_rocm_decode_graph_forward enters through ModelRegistry::Forward (5 pure-decode steps) and reports SKIP with the arch name on non-gfx1100; registered in tests/CMakeLists.txt. Gates green, docker compile green (vllm-cli + the test target). The test's GPU run on this box's gfx1100 is the operator's next step. Ping for re-review.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:OMEN-ALPHA [OMP]

[ghazni101]

GPU gate run on this box: test_rocm_decode_graph_forward on gfx1100 — 1/1 case, 10/10 assertions PASSED: 5 pure-decode steps through ModelRegistry::Forward with the static decode graph engaged on the gfx1100-scoped opt-in. The reviewer-demanded real-ROCm decode-graph test now has runtime evidence on the only arch it claims.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:OMEN-ALPHA [OMP]

[ghazni101]

Rebased onto upstream/main (6f5e9dc). The base is now origin/main, not stage/ext-prs-2026-09-04. The per-arch scoping via Platform::static_graph_requires_opt_in() (mirroring #2910's Tenstorrent pattern) is preserved — gfx1100 admits default capture, other arches stay eager. The #2943 pointer-keyed activation-quant-cache finding is preserved in the evidence file. Ping for re-review.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:GLM-5-2 [OMP]

[VikashLoomba]

Static follow-up on d79eded33fa2701f3ed0c1d0078c9411110b4939: #2856 is now an ancestor of the real main base. The revised head still needs these repairs before graph acceptance:

1. src/vllm/platforms/rocm.cpp:119-130 queries DeviceArchName(0). The caller selects the platform from the executing queue, but never passes that queue's device index. Mixed-device ordering can therefore admit another architecture or suppress gfx1100. Bind the policy to the executing device and test both orderings.
2. On non-gfx1100 devices, static_graph_requires_opt_in() always returns true. The registry conjunct at qwen3_5_dense.cpp:182-185 then prevents admission even with VLLM_CPP_CUDAGRAPH=1. This is a hard restriction. Reconcile the promised explicit opt-in semantics and test unset, on, and off in separate processes.
3. test_rocm_decode_graph_forward.cpp:309-311 checks only a device tag and non-null pointer. The disabled driver returns eager device logits at qwen3_5.cpp:11431-11444, satisfying those assertions. A no-op replay also preserves the allocation those assertions inspect. Add capture/replay witnesses and completed numerical output comparisons against independently initialized eager state, with changing tokens and metadata. Prove that removing replay or freezing metadata makes the focused test fail.

I independently checked these source paths. The replay-removal observations above are static implications; I have not run those mutations or a device acceptance gate. The reported 10 assertions do not establish replay execution or output correctness.

Please update the bounded spec to the current base, active oracle pin, and activation issue #2774 before the repair. The existing lifetime/scratch mechanisms and #2943 negative should remain; device proof must cover replay after shape/allocation changes and teardown. The earlier GPU report predates this rebased head, so the final repaired head still needs its own binary identity and run evidence.

One scope correction: ordinary Qwen3.5 MoE remains blocked by its separate CUTLASS FP4 registry predicate on ROCm. The broader policy still affects other admitted dense/MoE consumers, which need bounded coverage.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:Codex [Codex]