docs+manifest(rocm): close Still Open landed bullets + add compiled status for executing ops - #119
Conversation
…nded + new `compiled` status for executing ops Two reconciliations so the audit reflects ROCm execution reality. (1) ROCM_AUDIT.md — the `## Still Open` section had become a changelog: 6 of its 7 bullets described landed, executing, execute-compare-tested, runtime.launch()- reachable kernels. Renamed that heading to `## Landed — compiler-generated execution on gfx1151` and added a fresh `## Still Open` with only the genuinely- open work: the remaining artifact_only op surface (norms/activations/positional/ matmul-chains/exotic-attention), CDNA (hardware-gated), and the parked KV-tile pipelining rung. (2) backend_manifest — new `compiled` status (a rung below `hardware_verified`): "executes on hardware via runtime.launch() as a compiler-generated hsaco, with a checked-in execute_compare_fixture, but NO shipped C-ABI runtime_symbol." This is the honest status for the ROCm compiled-lane family that previously sat at `artifact_only` despite executing. The validator requires the fixture (not a symbol). Promotes 7 rocm rows: gqa_attention / mqa_attention / multi_head_attention / attn_sliding_window (flash_attn kernel), linear_attn / lightning_attention (linear-attn kernel), fused_epilogue (matmul kernel) — each pointed at its *_compiled.py execute-compare test. gpu_target_map renders + counts + legends the new rung; the stale "every row is artifact_only/planned" intro is corrected (NVIDIA sm_120 + ROCm now carry execution rungs). rocm_target_map: artifact_only 30→23, +7 compiled, 2 hardware_verified. 56 manifest/fixture tests pass; mypy + drift clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
gstoner
changed the title
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c009e89611
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| "gqa_attention": { | ||
| "dtypes": ("fp16", "bf16"), | ||
| "notes": "GQA/MQA via the flash_attn WMMA kernel (gqa directive attr; " | ||
| "fwd+bwd, grouped K/V; runtime detects from operand shapes). " | ||
| "Executes on gfx1151 via runtime.launch() (rocm_flash_attn_" | ||
| "compiled); no shipped C-ABI symbol.", | ||
| }, |
There was a problem hiding this comment.
For these ROCm flash-attn-family rows, the checked-in runtime path still only accepts tessera.flash_attn (runtime._execute_rocm_compiled_flash_attn, lines 1909-1913), and the GQA/sliding-window fixtures also stamp that op name rather than tessera.gqa_attention, tessera.mqa_attention, etc.; rocm_compiled similarly only accepts matmul/gemm for the fused epilogue. Any caller/audit consumer that tries to launch the manifest op names now marked compiled is rejected before GPU execution, so the target map overstates per-op runtime.launch() support unless the executors are taught those op names or these rows stay artifact-only.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Aliases now actually launch (not just marked compiled) — you were right that the executors rejected the alias op names; rather than revert those rows to artifact_only, I taught the executors the op names (consistent with how lightning/retention were handled in this same PR stack):
| counts[r.status] = counts.get(r.status, 0) + 1 | ||
| for status in ("fused", "compileable", "executable", "artifact_only", | ||
| "reference", "planned"): | ||
| for status in ("hardware_verified", "compiled", "fused", "compileable", |
There was a problem hiding this comment.
Adding compiled to the GPU map means all_rocm_rows('rocm') now yields that status, but the shared-vocabulary drift gate in tests/unit/test_gpu_target_maps.py still allowlists only through hardware_verified; running PYTHONPATH=python pytest -q tests/unit/test_gpu_target_maps.py::test_apple_and_gpu_dashboards_share_status_vocabulary fails with GPU dashboard uses unknown status(es): {'compiled'}. Please update the downstream status consumers/allowlist along with this renderer change so CI stays green.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Status-vocabulary drift gate — added compiled to the shared allowlist in test_gpu_target_maps.py::test_apple_and_gpu_dashboards_share_status_vocabulary (the rung below hardware_verified: runtime.launch() hsaco + numerical fixture, no shipped C symbol). CI stays green.
…ch executors the `compiled` op names
Two review points on the `compiled` promotion:
(1) Drift gate. Adding `compiled` to the GPU target map made
test_gpu_target_maps.py::test_apple_and_gpu_dashboards_share_status_vocabulary
fail ("GPU dashboard uses unknown status(es): {'compiled'}"). Added `compiled`
to the shared-vocabulary allowlist (the rung just below hardware_verified:
runtime.launch() hsaco + numerical fixture, no shipped C symbol).
(2) Don't mark aliases `compiled` before launch accepts them. The flash-attn
executor only accepted tessera.flash_attn and the gemm executor only
matmul/gemm, so the op names newly marked `compiled` (gqa/mqa/mha/sliding-window,
fused_epilogue) would have been rejected before GPU execution — the target map
would overstate runtime.launch() support. Taught the executors those op names
(matching how lightning/retention were handled):
* _execute_rocm_compiled_flash_attn accepts tessera.{flash_attn,
multi_head_attention,gqa_attention,mqa_attention,attn_sliding_window};
GQA/MQA come from operand shapes, window/softcap from kwargs.
attn_sliding_window additionally REQUIRES a positive `window` (the op name
implies one) — else a named error.
* rocm_compiled (gemm) accepts tessera.fused_epilogue (= matmul + the fused
bias/activation epilogue, same kernel).
Tests: GPU-free op-name acceptance for each alias (a downstream validation
fires, proving the op-name gate is passed, not the "handles exactly one"
rejection) + unknown-op rejection + attn_sliding_window-requires-window + a GPU
oracle launch of the multi_head_attention op name on gfx1151. So every op name
marked `compiled` genuinely launches via runtime.launch().
mypy + ruff + drift + manifest/fixture tests all clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
Both review points addressed in 1. Status-vocabulary drift gate — added 2. Aliases now actually launch (not just marked
Tests added: GPU-free op-name acceptance for each alias (asserts a downstream validation fires, proving the op-name gate is passed — not the "handles exactly one" rejection), unknown-op rejection, mypy + ruff + drift + 49 manifest/fixture tests clean. |
…_attention `compiled` promotion
CI fail: test_rocm_attention_family_uses_16x16x16 asserted
manifest_for("multi_head_attention")["rocm"].mfma_shape == (16,16,16,1), but
promoting multi_head_attention to `compiled` (WMMA, no MFMA shape — it IS the
flash_attn kernel) made that None.
The test's premise (MHA on the artifact MFMA path) is outdated. Repointed the
MFMA-shape assertion at a still-artifact_only attention op
(deepseek_sparse_attention, which keeps the 16x16x16 CDNA MFMA score-tile), and
added assertions that multi_head_attention is now `compiled` / wmma / no MFMA
shape — documenting the move off the artifact path. No hardcoded artifact_only
counts elsewhere; 166 manifest/coverage/conformance + 179 rocm execution tests
pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
CI fix pushed ( Fix: repointed the MFMA-shape assertion at a still- Checked there are no hardcoded |
1 participant
Recovery + the "close the Still Open list" task in one. This is commit
0f95a707, which was pushed to #118 but orphaned — #118 merged at the priortip (
f204ff7f, the op-name dispatch, which IS on main), and this audit-hygienecommit didn't land (third time a fast background merge captured an earlier tip).
Re-landed cleanly on current main. It also is the "close these landed bullets"
work that was requested.
What
1. ROCM_AUDIT
## Still Open→## Landed. That section had become achangelog — 6 of 7 bullets describe landed, executing, execute-compare-tested,
runtime.launch()-reachable kernels (matmul, flash_attn fwd+bwd, GQA/MQA,fused epilogue, sliding-window, logit-softcap, linear_attn incl.
lightning/retention). Renamed to
## Landed — compiler-generated execution on gfx1151; added a fresh## Still Openwith only the genuinely-open work(remaining
artifact_onlyop surface, CDNA hardware-gated, parked KV-tilepipelining).
2. New
compiledmanifest status — a rung belowhardware_verified:executes via
runtime.launch()as a compiler-generated hsaco + execute-comparefixture, but no shipped C-ABI
runtime_symbol. Promotes 7 rocm rows(gqa/mqa/mha/sliding-window → flash_attn kernel; linear_attn/lightning →
linear-attn kernel; fused_epilogue → matmul kernel), each tied to its
*_compiled.pytest.rocm_target_map: artifact_only 30→23, +7 compiled,2 hardware_verified. No inflation (Decision #25) —
compiled≠hardware_verified(no C symbol), explicit in the validator + legend.Verification
🤖 Generated with Claude Code