Re-derive gfx1100 decode-graph activation on current main with device and replay proof

[VikashLoomba]

Row: BACKEND-ROCM

Re-derive the gfx1100 decode-graph activation on current main. This issue supplies canonical implementation ownership for the unresolved work in #2774 and PR #2777. Their historical GFX1100-TG200 label is not a current roadmap/backend row. The original author owns those records; this issue preserves their provenance instead of rewriting their history.

At main f98b638673b4d2edc0250eec56d229357ea38ab1, the ROCm backend implements HIP graph capture/replay, but the platform leaves static graph admission false. The current #2777 head d79eded33fa2701f3ed0c1d0078c9411110b4939 is cumulative and its activation/test slice still has three independently inspected gaps:

• Admission queries physical device 0 rather than the executing queue's device index.
• The advertised explicit opt-in on other architectures remains a hard restriction.
• The production test checks device allocation, without completed output comparisons or a witness that graph replay executed.

The first deliverable is a committed bounded spec on current main, with active-vLLM pin e126687a9a828d513c01a07cd69f025f27d63280, exact device/model admission, explicit-on/off/default behavior, and a production correctness gate before activation acceptance. Scope must preserve existing shared graph, retained scratch, provider, dtype and sampling mechanisms. Keep the #2943 pointer-cache negative and all historical performance evidence, with their limitations.

The implementation must use the existing ModelRegistry/graph/backend path and prove capture and real replay. Its gate needs independently initialized eager/graph state, completed numerical outputs, changing token/position/block/GDN-slot metadata, shape/slot transitions, scratch growth, recapture and outstanding-work teardown within the admitted scope. A fresh reviewer must delete the production launch/replay call and freeze metadata to prove the tests detect those defects. Unsupported device or model modes must remain explicitly bounded rather than inheriting an untested broad activation.

Fresh implementation, mutation review and operator gates are required. The operator holds the local GPU mutex. Performance acceptance follows correctness and uses the same binary, matched controls and profiler, idle interleaved runs, source/binary/model identities, throughput, latency and memory. Historical contended numbers do not establish a current-main benefit. #2773 remains the owner of state/numerical characterization and its unresolved oracle checks.

The landing PR must link this issue and its committed spec. No merge or throughput acceptance is implied by filing it.

[VikashLoomba]

The current-main-derived candidate is c3c880d6abdfc6c698cee7e3bb1bbf4f73369062, based on main 2add10f31d28381a10c206596f62cd1878bf7205. Its approved-image gfx1100 build and scoped CPU mutation review are complete. The candidate is not yet ready to land.

Completed native controls:

• Qwen3.5-0.8B: 24 captures across auto, BF16, and FP8 KV; smoke and transition workloads; eager, ambient graphs, explicit graphs, and independent graph repeats. All 24 required comparisons match raw state, pre-sampling logits, tokens, and metadata exactly.
• Both gate models: 18 lifetime positives across the three KV modes and refresh, block-width change, and destruction. Pending device work completes before the production guard releases storage. The original GNU symbolization failure remains recorded; the reviewed ELF-bias calculation and approved-image LLVM symbolizer resolve the actual guard frames.
• Actual nonzero logical device: six 0.8B explicit-Queue runs completed on stream device 1. All three eager-versus-graph comparisons are exact. Every graph run captured once and replayed six times. This does not yet discharge physical PCI/trace attribution or the public selector gate.

A fresh data reviewer verified the first 42 capture/lifetime archives, their native input and binary bindings, comparisons, and nonoverlapping GPU-mutex intervals. The completed raw captures were archived, verified, and removed from scratch. The six nonzero-device controls retain a separate operator report and await fresh data review.

The 4B quantized reader repair #3056 is frozen at 23b23f56322e8434eb7621e127dacd751dc8abb7 for fresh review. It distinguishes the actual split quantized MLP provider chain from the merged BF16 chain. Product C++ and the hardware binary remain unchanged. Initial 4B device-1 eager and graph runs both complete; state acceptance waits for that reader review and the operator gate.

Hardware fault controls remain required. The private preparation helper for deleting the actual HIP graph launch and the two graph-cleanup calls is frozen for independent review. Its passing CPU/archive tests are not hardware mutation evidence. Other changing-input, scratch, padding, ordinary C API/CLI, oracle/default, and performance gates remain open.

The native model trace reproduces the invalid UTF-8 exporter defect in #3051. Partial CSV graph-launch joins and an exact profiled/unprofiled state comparison do not repair the full trace gate. The numerical prerequisite in #2773 also remains failing. No new throughput claim or default change is accepted.

[VikashLoomba]

The 4B quantized reader passed fresh review and operator gates. All 24 ordinary 4B captures passed exact state, logits, token, and metadata comparisons across auto, BF16, and FP8 KV modes. Independent data review also verified the nonzero-device and ordinary CLI controls. These results extend the earlier 0.8B evidence without changing product head c3c880d6abdfc6c698cee7e3bb1bbf4f73369062.

The 36 graph-cleanup fault controls were repeated with retained commands and kernel-visible mutex ownership before and after every execution: 12 intended failures and 24 controls passed their expected results. Deleting the actual HIP graph launch changes the first affected token. Its isolated detector-removal pair remains incomplete because other exact-state checks also reject the same output; that limitation remains recorded.

Diagnostic head 454fbddfc508f1a9af910ee5809ec94671ccf001 adds larger prefill followed by older-graph reuse, survivor permutation, block-width and cache-backing changes, padded-slot readbacks, and executing quant-scratch observations. The expanded native matrix is in progress. The full repository preflight and inherited reader gates passed. Fresh review found an exception-path ownership leak in the diagnostic after successful device allocation but failed host bookkeeping. A fresh implementer is repairing it before final acceptance.

The remaining requirements include the repaired diagnostic's review, native defect controls, complete physical trace attribution (#3051), active-oracle control alignment (#3046/#3061), and the unchanged numerical failures (#2773). No new throughput result or graph-default change is accepted. Completed scratch captures and retired worktrees are archived and verified before removal.