feat(KERNEL-QUANT-CIQ-GEMM-ROCM-RDNA3): enable gfx1100 WMMA prefill - #3187
Open
VikashLoomba wants to merge 9 commits into
Open
VikashLoomba wants to merge 9 commits into
VikashLoomba wants to merge 9 commits into
Conversation
The generic rocWMMA tile API already supplies RDNA3 operand transforms. Prove that admitting gfx1100 reuses the Q4_K and Q6_K bodies before changing arithmetic. Keep attention admission separate and compare fresh upstream snapshots requested by the developer. Tracks ISSUE-LOCAL-01M2F0PQWGSCXG0N4951NF9DPZ. Staged preflight matches the pristine baseline: test_agent_onboard assumes master although global Git config selects main. Its 39 tests pass with GIT_CONFIG_GLOBAL=/dev/null. Twelve unrelated gates report explicit skips. Record and staged-diff checks pass. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:gpt-6-astra [Codex]
…prefill Reuse the existing Q4_K and Q6_K rocWMMA tiles on gfx1100. The fragment API already resolves its operand layout. Keep attention admission separate. Hardware red-first tests fail ten dispatch assertions before admission. The unchanged kernels pass all four focused cases and 46 assertions after admission. The scalar control passes the same cases in a separate process. The public completion gate reaches both formats and matches all 1024 logits and eight generated tokens against its scalar child process. Port all 240 original plugin test_mmq Q4_K/Q6_K tensor cases, preserving inputs and tolerances. Native outputs pass against both the plugin and its dense reference. F16 output narrows from native F32 in the test adapter. The gfx1100 code object contains signed-int8 WMMA instructions. This is an immutable local implementation checkpoint. The staged preflight is running. Broad HIP gates, mutation review, and full-model performance remain pending. No performance claim or remote handoff is authorized by this checkpoint. Tracks ISSUE-LOCAL-01M2F0PQWGSCXG0N4951NF9DPZ. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Codex:GPT-6 [Codex]
…receipts Keep the physical red-first results, original primary fixture comparisons, public scalar control, emitted instructions, and negative mutations with the architecture admission change. Both executing tile bodies retain their committed-spec hashes. The complete HIP gate passes all 29 registered tests with five unchanged resource skips. The full staged preflight exits zero with 12 unchanged skips. Four CPU and four public production mutations fail as intended, with immutable source hashes preserved. Full-model performance and fresh review remain with the coordinating operator. Tracks ISSUE-LOCAL-01M2F0PQWGSCXG0N4951NF9DPZ. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Codex:GPT-6 [Codex]
…ication Preserve operator and fresh-review evidence for the unchanged implementation. The review passes with no findings. The operator HIP gate passes with the same five resource skips, and all seven supplemental NumPy suites execute. Keep the original harness failures and unresolved model gate explicit. Tracks ISSUE-LOCAL-01M2F0PQWGSCXG0N4951NF9DPZ. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Codex:GPT-6 [Codex]
…ssion Record exact model input and output identity, interleaved native results, prefill-only WMMA traces, and the operator's 240 original matrix cases. Close architecture admission with the reviewed implementation unchanged. Keep every full-model floor and clock-attribution obligation open under ISSUE-LOCAL-01M2F4WCD6ZK5VH5S8TF83APD6. Flatten evidence into the classifier's permitted per-run layout. Preserve source bytes and hashes, update copy maps, and reconcile only the parent's gfx1100 record. No product, test, tool, or checker changes are included. The full preflight exits 1 on the known onboarding fixture's branch-name assumption and retains 12 skips. Its unchanged 39-case suite passes with GIT_CONFIG_GLOBAL=/dev/null. Scoped record, link, hash, role, and staged checks pass. Prior independent implementation review remains PASS. Closes ISSUE-LOCAL-01M2F0PQWGSCXG0N4951NF9DPZ. Closes ISSUE-LOCAL-01M2F51K16TDAC95DM1VEKBBCK. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Codex:GPT-6 [Codex]
VikashLoomba
marked this pull request as ready for review
September 14, 2026 05:29
The developer requested a reviewable diff with the raw evidence preserved outside the pull request. ISSUE-LOCAL-01M2GNY58NHBK3D4JQQ738M6GR defines lossless archival, public download verification, and an unchanged product. The exact d6e source already passed the recorded local preflight checks, with five explicit argument-dependent skips. The new role and scoped canonical-record checks pass before this documentation change. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:gpt-6-astra [Codex]
…t repair Bring the reviewed upstream keep-quant fixture repair into the row branch before publishing the evidence packaging correction. ISSUE-mudlerGH-3188 owns the resolved buffer overrun and NaN comparison. All nine WMMA product, test, and harness files remain identical to the reviewed d6e head. Imported records match their complete origin/main versions. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:gpt-6-astra [Codex]
The pull request carries 236 raw evidence files for architecture admission. Keep one report and preserve every original byte in the published archive. Archive and per-file checksums pin the evidence to its source revision. The report retains correctness results and all performance qualifications. Implementation, tests, and validation harness bytes remain unchanged. Closes ISSUE-LOCAL-01M2GNY58NHBK3D4JQQ738M6GR. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Codex:GPT-6 [Codex]
… directory The retrieval subshell discarded its temporary destination without reporting where readers could find the extracted reports. Print the absolute directory after all 236 original file checksums pass so readers can open those reports. Archive verification still precedes directory creation and extraction. Repairs the retrieval recipe for ISSUE-LOCAL-01M2GNY58NHBK3D4JQQ738M6GR. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Codex:GPT-6 [Codex]
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Enable Q4_K and Q6_K WMMA prefill on gfx1100 through the existing quantized matmul dispatch. Both tile bodies remain unchanged. Attention keeps its separate architecture gate, and
VT_ROCM_QUANT_WMMA=0keeps the scalar control.Scope
Row
KERNEL-QUANT-CIQ-GEMM-ROCM-RDNA3. ResolvesISSUE-LOCAL-01M2F0PQWGSCXG0N4951NF9DPZand evidence-packaging issueISSUE-LOCAL-01M2GNY58NHBK3D4JQQ738M6GR. The developer selected one PR. Specification commita5b5c92precedes implementationc3fe98ba6.The runtime change touches two files. The review diff retains tests, the comparison harness, required records, and one evidence report. All 236 original evidence files are preserved in the downloadable archive, pinned to source
d6e40c91f634a041c873c7a04516d55c4d05772a. The retained report records its checksum and extraction procedure.Validation
Exact revisions, model identity, commands, review receipts, failed attempts, and raw results are retained in the archive. The summary is
docs/bench-evidence/rocm-rdna3-quant-wmma/README.md.Qualifications
Local strict readiness retains five argument-dependent preflight skips. PR classification passes separately; the remaining checks concern unchanged ARM/CPU/CUDA/Triton paths. The original CPU CI failures came from the keep-quant reference fixture and were repaired upstream in PR 3189, now included in the branch.
Full-model performance floors remain unresolved under
ISSUE-LOCAL-01M2F4WCD6ZK5VH5S8TF83APD6. Dynamic clocks, unequal cache capacities, and instrumentation differences prevent accepted speedup or memory-efficiency claims. Token parity applies to the retained workload. Other gfx11 devices remain excluded, and RDNA4 hardware was not rerun.FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:gpt-6-astra [Codex]