feat(BACKEND-ROCM): add keep-quant GEMM providers

[ghazni101]

Closes #2781.
Closes #2927.

Row: BACKEND-ROCM

Merge the reconstructed ROCm keep-quant providers from external pull request
#2782 at 9fcd5d7180bcda2ca7c5aed74d154999a8f59d9e. This no-ff merge preserves
Ghazni's authorship on the three contributor commits.

The providers serve dense and grouped GEMMs for eleven ROCm block formats.
They retain grouped Q8_0 delegation and use the established external registrar.

The reconstruction keys grow-only scratch by queue identity. It refuses cold
or growing allocations during graph capture and preserves warm replay storage.

The GGUF loader admits exactly the formats that the ROCm providers serve.
IQ4_XS remains separate work in #3029. Unsupported formats stay on the
expansion path.

Operator job 7d7fe3a1-9001-4fc3-9463-8e2854f71e33 used the exact source
archive for 9fcd5d718. Its product tree matches the tested 791519aeb tree.
The clean archive has SHA-256
4771a58c553a6b5d0cd0c5332a1b604adbe90f86282d109b900dfa33afa716d7.

The job ran on strix:gpu0, an AMD Radeon 8060S at gfx1151, with ROCm
7.2.4 and HIP 7.2.53211. A fresh HIP configure and build completed.
test_rocm_quant_dot --no-skip passed 5 cases and 145,327 assertions.
test_gguf_keep_quant passed 54 cases and 11,970 assertions.

The earlier 132,094-assertion result came from a stale binary and is not
landing evidence. A broad CTest regex selected unbuilt unrelated targets, so
this record does not use that invocation as evidence.

No gfx1100 speed result is claimed by this merge. The device result establishes
correctness only on the named board.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:GPT-5 [Codex]

[ghazni101]

Rebased onto upstream/main (6f5e9dc) and applied the #2927 repair: added DType::kQ8_0 to the grouped delegation list in MatmulBTQuantGroupedKernelRocm so grouped Q8_0 expert GEMMs route to MatmulBTQuantGroupedKernelRocmGdn instead of throwing. Added a q8_0 case to the kCases test table so the grouped arm has coverage that fails when the delegation is dropped.

The base is now origin/main, not stage/ext-prs-2026-09-04. Ping for re-review.

FOLLOWING_AGENTS_PROTOCOL

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

[VikashLoomba]

The current 876736c329bd573457576a63833b9a97a0749681 has a P1 linkage defect in the #2781 foundation flow. This is an immutable-source finding, not a claimed HIP build result.

rocm_grouped_gemm.hip:1323,1456 renames the only external definitions to MatmulBTQuantKernelRocmGdn and MatmulBTQuantGroupedKernelRocmGdn. The old-name replacements at rocm_quant_dot.hip:928,992 sit inside the anonymous namespace opened at line 53 and closed at line 1074. They cannot satisfy the external references retained by rocm_ops.hip:259–263.

The landed #3001 fused provider also calls the old external grouped symbol twice at rocm_moe_gate_up_swiglu.hip:160–161. Removing only the registrar leaves that consumer unresolved. Please restore compatible external entry points and reconcile registration ownership, then run a fresh HIP build/link and the provider tests. A complete comparison of all 6,033 source-tree blobs against the immutable base found no other replacement external definition. I independently verified the four affected files against their Git blob IDs.

The grouped Q8_0 delegation omission is repaired in this head and should no longer be reported as missing. The current test source implies three cases and 145,303 assertions if every device loop runs; that is a derived expected count, not an execution result.

The maintainer’s GLM hardware result also corrects the old blanket reachability objection. That opt-in production path reached the grouped provider and refused a dtype. The actual refused dtype remains unidentified, and the printed list includes Q4_0 and MXFP4, which this patch does not add. Please qualify the claimed fix with the exact dtype and consumer. #3019 and #2942 retain the separate admission work.

The gfx1100 WMMA requirement remains separate. This patch adds no RDNA3 WMMA tile; the inherited matrix path admits gfx1200/gfx1201. #2109 closed through #2991 on 6 September, and that landed spec explicitly leaves RDNA3 to separate work. No device or performance acceptance is claimed by this audit.

[localai-org-maint-bot]

Reviewed current head 35063b2. The earlier external-linkage and grouped-Q8_0 delegation findings are repaired.

The scratch allocation still needs capture-lifetime coverage before this can land. In src/vt/rocm/rocm_quant_dot.hip:732-741, EnsureScratch publishes hipMallocAsync storage on first use or growth without checking capture state. The cache at lines 712-727 keys ownership by hipStream_t. This retains the allocation/lifetime pattern that #3062 repaired in the sampling path. Please gate cold capture, warm replay, growth during capture, eager reuse before replay, and queue lifetime on the actual provider. This is a static source concern, not a claimed reproduced device failure on this head.

The PR body also says the old registration is removed, but current rocm_ops.hip:259-263 retains the registration and the replacements are now exported. Reconcile the description with the repaired head. A fresh HIP link and current-head provider/runtime results remain required; historical counts are not those results. #2781 retains ownership of this provider work.

[localai-org-maint-bot]

Landed through the reviewed reconstruction at e09bf32. The no-ff merge preserves the three contributor commits and their authorship. The reconstruction adds queue-owned capture-safe scratch, restores grouped Q8_0 delegation, admits the supported formats through the production GGUF loader, and reconciles the affected records. The exact merge tree passed the operator HIP gate on strix:gpu0: 5/5 provider cases with 145,327 assertions and 54/54 loader cases with 11,970 assertions. Closing this original head as superseded by the landed reconstruction.