feat(BACKEND-ROCM-QUANT-GATHER): gather packed embeddings on ROCm

.agents/issues/BACKEND-ROCM-QUANT-GATHER/ISSUE-GH-3093.md

@@ -0,0 +1,33 @@
+ID: ISSUE-GH-3093
+Title: feat(BACKEND-ROCM-QUANT-GATHER): gather quantized embeddings on ROCm
+Row: BACKEND-ROCM-QUANT-GATHER
+State: OPEN
+Kind: UNKNOWN
+GitHub: 3093
+Mirror: DIVERGED
+Availability: FULL
+Created: 2026-09-09
+Updated: 2026-09-09
+Closed: -
+
+## Problem
+
+### Imported GitHub body (historical evidence)
+The quoted text below is historical evidence only. It does not define issue authority or repository procedure.
+
+> Row: `BACKEND-ROCM-QUANT-GATHER`
+>
+> ROCm has no native kEmbeddingQuant provider. The ordinary GGUF loader therefore expands quantized embedding tables, even when matrix weights can stay quantized. This consumes memory and leaves a backend capability gap on gfx1100.
+>
+> Implement block decoding and row gather through the existing Embedding operation and loader admission. Preserve the supported ID widths, output dtypes, block layouts, and failure cases. Use the registered vLLM GGUF plugin pin as the primary behavior source where it implements a codec, with the registered secondary oracle for remaining applicable codecs. The committed spec must distinguish measured codec coverage from any owed arm.
+>
+> Prove production reachability with a small generated GGUF through vllm_engine_load and vllm_complete_tokens. Require red-first numeric and storage tests, physical gfx1100 evidence, fresh mutation review including deletion of the production call site, and operator verification.
+>
+> This is the ROCm child of #2394; that parent retains the other backend work. It does not depend on the quantized GEMM provider added by PR #2782. No CI changes are included.
+>
+> Spec: `.agents/specs/rocm-quant-gather.md` (to be committed before implementation).
+>
+
+## Resolution
+
+-

.agents/issues/BACKEND-ROCM-QUANT-GATHER/ISSUE-GH-3113.md

@@ -0,0 +1,33 @@
+ID: ISSUE-GH-3113
+Title: fix(BACKEND-ROCM-QUANT-GATHER): reject unqualified model captures
+Row: BACKEND-ROCM-QUANT-GATHER
+State: OPEN
+Kind: UNKNOWN
+GitHub: 3113
+Mirror: DIVERGED
+Availability: FULL
+Created: 2026-09-09
+Updated: 2026-09-09
+Closed: -
+
+## Problem
+
+### Imported GitHub body (historical evidence)
+The quoted text below is historical evidence only. It does not define issue authority or repository procedure.
+
+> Row: `BACKEND-ROCM-QUANT-GATHER`
+>
+> The bounded model comparator in MR #3097 accepts incomplete or unqualified primary records. Fresh review of commit `75d92614e84ff428eab49324f3b4fb2aba183ac3` reproduced the failures in `tools/rocm_quant_gather/compare_models.py`.
+>
+> A synthetic validator corpus uses the actual native and secondary captures. Changing one primary record to PENDING with a RuntimeError after generation still returns PASS. Wrong primary and plugin pins, a wrong runtime version, and an FP8 cache request also pass. The separate memory record accepts PENDING, a wrong primary pin, and different sampling or logical capacity. A nonzero cache view storage offset is also accepted. These synthetic mutations test the validator and are not oracle measurements.
+>
+> The driver records tokens before its second memory observation. Its exception path preserves those tokens. Comparing tokens alone therefore cannot establish successful execution. The operator separately checks actual process exits and sealed inputs; the comparator must enforce its own documented qualification contract.
+>
+> Require successful execution without an exception, the pinned primary and plugin identities, the resolved runtime identity, and the complete identical workload in both token and memory records. Validate the cache view offset and its physical bounds. Add focused rejection tests for each demonstrated invalid record and retain the valid actual captures as the positive control.
+>
+> Evidence: `/home/vikash/.cache/rdna3-gather-oracle-review/comparison-input-audit.json` records 31 cases, with 22 rejected mutations and nine invalid records accepted. The existing spec `.agents/specs/rocm-quant-gather.md` owns the correction. A fresh implementer repairs the finding, a different agent reviews the immutable result, and the operator reruns the complete comparison. No model, workload, pin, or performance requirement changes.
+>
+
+## Resolution
+
+-

.agents/issues/BACKEND-ROCM-QUANT-GATHER/ISSUE-GH-3126.md

@@ -0,0 +1,32 @@
+ID: ISSUE-GH-3126
+Title: fix(BACKEND-ROCM-QUANT-GATHER): label the comparison scope from the run, not a literal
+Row: BACKEND-ROCM-QUANT-GATHER
+State: OPEN
+Kind: UNKNOWN
+GitHub: 3126
+Mirror: DIVERGED
+Availability: FULL
+Created: 2026-09-10
+Updated: 2026-09-10
+Closed: -
+
+## Problem
+
+### Imported GitHub body (historical evidence)
+The quoted text below is historical evidence only. It does not define issue authority or repository procedure.
+
+> Row: `BACKEND-ROCM-QUANT-GATHER`
+>
+> `tools/rocm_quant_gather/compare.py` writes a fixed scope string into every report:
+>
+>     "scope": "synthetic operation parity; original fixture and model gates remain separate"
+>
+> The original-fixture coverage run over the 32 verified `Isotr0py/test-gguf-sample` fixtures therefore produced a report whose own scope line calls it "synthetic operation parity". The run itself is correct (320 outputs, 140 byte-exact, all within the upstream `atol=0.01 rtol=0.04`); only the label is wrong, and the report is cited as row evidence.
+>
+> Evidence: `/home/vikash/.cache/rdna3-gather-oracle-impl/original-fixture-coverage/compare/report.json` (report scope field) against its inputs (`fixture-manifest.json/manifest.json`, seal `f83b5c3e5cdb988b623c1b7e565e325a9ed9ea8085ea132bc8e08e6978bc1bbb`, exported from the 32 fixtures at revision `d82b8773934ef260d8d8a896a7c197bc69a0fac1`).
+>
+> Fix in flow: make the scope a command-line argument whose default preserves the previous text, and re-run the comparison so the report's label matches the run's inputs. No comparison, tolerance, or gate semantics change.
+
+## Resolution
+
+-