[2/5] autotune: two-track config + first real adopter (rmsnorm) (#770) by jhinpan · Pull Request #785 · ROCm/FlyDSL

jhinpan · 2026-07-01T07:05:40Z

Second in the #770 series: give the autotuner an avoid-search path and put it to work on a real kernel, so it stops being dead code.

Stacked on #783 (1/5). Targets main (cross-fork PRs can't base on a fork branch), so it currently shows PR1's commit too — review only eb94a0ac, or merge #783 first and this will shrink to its own diff.

Builder mode

Every current FlyDSL kernel bakes its structural knobs at module-build time — they size shared storage, the vectorized tile stride, and the launch block, and known_block_size is fixed at @kernel decoration — rather than exposing them as jit Constexpr params. The existing @autotune, which injects config kwargs into a single jit call, can't tune those.

So this adds builder mode: build_fn(config, *args, **kwargs) -> launch_callable rebuilds the module per config; built modules are cached per (key, config). Structural config kwargs route to build_fn (not the launch call); build-only kwargs like dtype_str are filtered out of the launch call by signature.

Two-track config

== Triton @heuristics + @autotune, quack get_default + exhaustive. New default= heuristic on @autotune: normal runs take the analytic default and skip the search entirely (zero-search); set FLYDSL_AUTOTUNE=1 to force the full sweep. Autotuners without a default always search.

First adopter: rmsnorm

build_rmsnorm_module gains a BLOCK_THREADS build knob (adds known_block_size when >256); the default arg keeps the old signature working, so the existing rmsnorm test path is unchanged.
kernels/rmsnorm_config.py — get_default (analytic BLOCK_THREADS) + get_all_configs (BLOCK_THREADS × waves_per_eu, fast-path only). VEC_WIDTH is intentionally not tuned (pinned to 128/elem_bits by the 128-bit buffer copy).
kernels/rmsnorm_autotune.py — rmsnorm_autotuned, the tuned front end.

Verification (MI350X / gfx950, M=4096 N=8192 bf16)

Default and forced-search runs both match the torch reference (max err 1.5e-2). The 12-config sweep runs cleanly and picks BLOCK_THREADS 256–512; the result is cached (second call doesn't re-tune):

[autotune] tuning 12 configs...
  [4/12] Config(BLOCK_THREADS=256) -> 0.081 ms
  [7/12] Config(BLOCK_THREADS=512) -> 0.081 ms
  ...
[autotune] best: Config(BLOCK_THREADS=256) (0.081 ms)

Tests

+6 GPU-free unit tests (builder rebuild/cache, default skip/force, kwarg filtering, env gate) → 22 total in tests/unit/test_autotune.py.
tests/kernels/test_rmsnorm_autotune.py (l2_device): default run, forced-search + tuned-config persisted, cache reuse.

ruff + black clean.

Refs #770.

🤖 Generated with Claude Code

Copilot

Pull request overview

This PR extends FlyDSL’s autotuning system to support a two-track “default heuristic vs forced search” flow and introduces a first real adopter (RMSNorm) that exercises builder-mode autotuning (rebuilding modules per structural config).

Changes:

Add builder-mode support to Autotuner (build_fn-based rebuild + in-process build cache) and introduce a heuristic default= path that skips search unless FLYDSL_AUTOTUNE=1.
Extend autotune machinery with stride normalization + extra cache-key axes, and add unit tests covering builder/default behavior.
Add RMSNorm autotune integration (rmsnorm_config.py + rmsnorm_autotune.py) and a GPU integration test.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
`python/flydsl/autotune.py`	Adds builder-mode execution, two-track default/forced-search behavior, and additional cache key axes; updates benchmark/run path accordingly.
`kernels/rmsnorm_kernel.py`	Adds a build-time `BLOCK_THREADS` knob and applies `known_block_size` when needed.
`kernels/rmsnorm_config.py`	Introduces heuristic default + exhaustive config generation for RMSNorm tuning.
`kernels/rmsnorm_autotune.py`	Provides the tuned RMSNorm front-end using builder-mode autotuning.
`tests/unit/test_autotune.py`	Adds GPU-free unit tests for builder mode, default skip/force behavior, kwarg filtering, and env gating.
`tests/kernels/test_rmsnorm_autotune.py`	Adds an end-to-end GPU test validating default correctness, forced search, persistence, and cache reuse.

Comments suppressed due to low confidence (1)

python/flydsl/autotune.py:250

In builder mode (fn=None), the disk-cache filename falls back to "unknown.json", so different tuners in the same cache dir can clobber each other’s tuned results. Consider deriving the cache filename from fn when present, otherwise from build_fn (and include the module name) so each tuner has an isolated cache file.

        # Disk cache
        fn_name = getattr(fn, "__name__", None) or getattr(fn, "func", None)
        if fn_name is not None and not isinstance(fn_name, str):
            fn_name = getattr(fn_name, "__name__", "unknown")
        fn_name = fn_name or "unknown"

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

+        # Dtypes + normalized strides of tensor args for type/layout specialization
        dtype_parts = []
+        stride_parts = []
        for name, val in sig_args.items():
            if hasattr(val, "dtype"):
                dtype_parts.append(f"{name}:{val.dtype}")
+            if hasattr(val, "shape") and hasattr(val, "stride"):
+                stride_parts.append(f"{name}:{_normalize_strides(val)}")
        key_vals.append(tuple(dtype_parts))
+        key_vals.append(tuple(stride_parts))


        def kernel_call():
            if config.pre_hook:
                config.pre_hook(merged_kwargs)
+            self._restore_tensors(snapshot)
            self._reset_tensors(args, merged_kwargs)


        compiler_opts = config.compiler_opts()
+        fn = self._resolve_fn(config, key, args, kwargs)
+        merged_kwargs = dict(self._filter_call_kwargs(fn, kwargs))
+        # In builder mode the config's structural kwargs (e.g. BLOCK_THREADS)
+        # are consumed by build_fn, not passed to the launch call.
+        if self.build_fn is None:
+            merged_kwargs.update(config.all_kwargs())
+


+        """Run a single (already-chosen) config: resolve fn, apply kwargs+hints."""
+        fn = self._resolve_fn(config, key, args, kwargs)
+        merged = dict(self._filter_call_kwargs(fn, kwargs))
+        if self.build_fn is None:
+            merged.update(config.all_kwargs())
+        return self._run_with_hints(fn, config.compiler_opts(), args, merged)


FlyDSL's autotuner exists but nothing uses it, and two gaps block real adoption. This is the first of a series making it a correct, adopted path. Cache key (_make_key) previously specialized on shape/dtype only. A config tuned under one compiler build, GPU arch, or memory layout would be silently reused under another. Fold in the axes Triton/quack rely on: - normalized stride pattern ({0,1,other}: broadcast vs contiguous vs strided) - device arch (get_rocm_arch) - toolchain fingerprint (reuses jit_function._flydsl_key) - cache-invalidating env vars (reuses _cache_invalidating_env_values) The dtype/stride axes are sorted by arg name so a call is keyed identically regardless of kwarg order (no duplicate tuning / cache files). restore_value (new) is the correctness soul of autotune: benchmarking runs the same kernel dozens of times, so an in-place / accumulating kernel (e.g. fused-add rmsnorm) corrupts its own inputs and picks a config on garbage. Snapshot the named tensors once and restore before every rep. reset_to_zero is now also re-applied on the real (non-benchmark) call — both the post-tune run and cache hits — via a shared _run_config, so an accumulate-into-zero kernel returns the single-clean-run result instead of carrying benchmark-rep state. (Was applied only inside the bench loop.) Also defer the CompilationContext import so the autotuner core stays importable and unit-testable without the compiled flydsl._mlir bindings. Adds tests/unit/test_autotune.py: 19 GPU-free tests covering Config serialization, every cache-key axis (incl. env-fingerprint change and kwarg-order insensitivity), restore_value/reset_to_zero semantics (incl. the final-run and cache-hit reset), pruning, and disk-cache round-trip. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Comment-only cleanup of the PR1 additions: keep the one key fact per helper, drop the Triton/quack background, redundant restatements, and by-example prose. No logic change; 19 unit tests still pass. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

@autotune

Second in the series. Gives the autotuner an "avoid-search" path and puts it to work on a real kernel, so it stops being dead code. Builder mode. Every current FlyDSL kernel bakes its structural knobs at module-build time rather than exposing them as jit Constexpr params. The existing @autotune, which injects config kwargs into one jit call, can't tune those. Add builder mode: build_fn(config, *args) -> launch_callable rebuilds the module per config. autotune_builder(): one-call adoption. Instead of a hand-rolled wrapper per kernel, a kernel opts in with just its kernel-specific pieces: rmsnorm_autotuned = autotune_builder( name="rmsnorm", build=build_rmsnorm_module, specialize=lambda inp, g, out, m, dtype_str="bf16", **kw: { "N": inp.shape[-1], "dtype_str": dtype_str}, configs=get_all_configs, default=get_default, structural=("BLOCK_THREADS",)) The helper owns cache naming, callable config generation, the structural-vs- compiler knob split, build caching, and launch-kwarg filtering. rmsnorm's adopter dropped from ~79 lines of boilerplate to a single declaration. Correctness (surfaced by two independent fresh reviews): - Build cache keys only on the STRUCTURAL knobs + spec, not repr(config), so configs differing only in a compiler hint (waves_per_eu) reuse one built module. Verified on MI350X: a 12-config sweep now builds 4 modules, not 12. - A build-only scalar (dtype_str) passed positionally is rejected with a clear error instead of silently binding to the wrong launch slot (e.g. stream). - autotune_builder requires a non-empty name (empty/None would fall back to unknown.json and defeat the per-kernel cache identity). - Build-only scalars enter the cache key via the specialize() axes. - Compiler hints (waves_per_eu / maxnreg) are folded into the JitFunction's compile_hints (restored after) so each distinct hint compiles a distinct binary instead of reusing a cached one. - FLYDSL_AUTOTUNE=1 bypasses cache + default to force a fresh search. - Disk cache is re-loaded when FLYDSL_AUTOTUNE_CACHE_DIR changes (a module- level tuner isn't pinned to the import-time dir for loads or saves). - Config.pre_hook is documented as non-persistable (not serialized). Two-track config == Triton @Heuristics + @autotune: default= gives zero-search normal runs; FLYDSL_AUTOTUNE=1 forces the sweep. First adopter rmsnorm: build_rmsnorm_module gains a BLOCK_THREADS build knob (+known_block_size when >256; default arg keeps the old signature). config space in rmsnorm_config.py; small-N (imported SMALL_N_THRESHOLD) emits a single config since that kernel ignores BLOCK_THREADS. VEC_WIDTH stays pinned (128b copy). Verified on MI350X (gfx950), M=4096 N=8192 bf16: default and forced-search match the torch reference (max err 1.5e-2); sweep picks BLOCK_THREADS 256-512 and a subsequent normal call reuses the cache (zero benchmark invocations asserted). Tests: GPU-free unit tests for builder mode + autotune_builder (rebuild/cache, build-cache-ignores-hints, default skip/force, scalar-in-key, name required, positional-scalar rejected) = 31; tests/kernels/test_rmsnorm_autotune.py (l2_device) covers default, forced-search, and no-re-tune cache reuse. ruff + black clean. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The autotuner routes Config(waves_per_eu=...) through compile_hints -> --amdgpu-waves-per-eu on gpu-module-to-binary opts=, which the AMDGPU backend silently ignores (a known limitation). So the rmsnorm _WAVES_PER_EU_CHOICES sweep tuned nothing. Lower the waves_per_eu compile-hint onto the kernel's rocdl.waves_per_eu gpu.func attribute in MlirCompiler.compile (before convert-gpu-to-rocdl) -- the same path the MoE kernels use. convert-gpu-to-rocdl turns it into the LLVM amdgpu-waves-per-eu function attribute, which the backend honors. Verified on MI355X (gfx950): the compiled LLVM IR now carries "amdgpu-waves-per-eu"="N" (previously absent) and the backend acts on it (ISA changes; occupancy warning at aggressive values). The JIT compile cache key already includes compile_hints, so each waves_per_eu recompiles a distinct binary. Co-authored-by: Cursor <cursoragent@cursor.com>

Follow-up to the waves_per_eu fix: the same "silently dropped compile hint" bug hit maxnreg, and num_warps had a related builder-mode trap. - maxnreg: Config(maxnreg=...) reached codegen only via --amdgpu-num-vgpr on gpu-module-to-binary opts=, which the AMDGPU backend ignores. Unlike waves_per_eu, amdgpu-num-vgpr has no native ROCDL translation, so lower it onto the gpu.func LLVM `passthrough`; convert-gpu-to-rocdl copies that to the llvm.func where the emitter turns it into the amdgpu-num-vgpr function attribute (same approach as FlyROCDLClusterAttrPass for amdgpu-cluster-dims). Generalize _apply_occupancy_compile_hints to a hint -> func-attr lowering. Verified on MI355X: the IR now carries "amdgpu-num-vgpr"="N" and the backend acts on it (ISA changes when the cap binds, e.g. 40 < the kernel's 60 VGPRs; 64/128 don't bind); outputs stay correct. - num_warps: in builder mode the block size is baked into build_fn, so a jit-kwarg num_warps can't be routed to the launch call and was silently dropped. Reject it with a clear error instead of tuning a no-op. - Tests: assert the func-attr lowering and the builder-mode num_warps rejection. Co-authored-by: Cursor <cursoragent@cursor.com>

…cache, f32) - Thread-safety: _run_with_hints no longer mutates the shared, cached fn.compile_hints (a race across concurrent tuned/served calls); it sets only the thread-local CompilationContext, and JitFunction folds the thread-local hints into its cache key + compile push, so each distinct waves_per_eu/maxnreg still compiles a distinct binary. - Search loop no longer silently swallows failures: unroutable configs (num_warps in builder mode) are rejected up front so they fail loudly even in a forced sweep, and "all configs failed" chains the last underlying error. - Disk cache: atomic write (tmp+rename); a single malformed entry is skipped rather than discarding the whole cache; a stale cached entry degrades to the default instead of crashing a normal call. - Cache key folds a source fingerprint of the adopter build/config functions, so editing the kernel/config invalidates a stale tuned best. - maxnreg lowering (_set_passthrough) replaces an existing amdgpu-num-vgpr entry instead of appending a duplicate. - restore_value/reset_to_zero run as an untimed per-rep setup (do_bench gains a setup param), so the restore copy no longer inflates the measurement. - get_all_configs sweeps f32 too (scalar path is BLOCK_THREADS-strided; it was silently collapsing to the single default). Verified on MI355X: f32 configs build and match reference (max_err 2.9e-6). - Builder mode rejects config kwargs not in structural (would route nowhere). Adds unit tests for each behavior. Co-authored-by: Cursor <cursoragent@cursor.com>

jhinpan · 2026-07-07T06:26:20Z

/rerun-ci

…#783/ROCm#807) Co-authored-by: Cursor <cursoragent@cursor.com> # Conflicts: # python/flydsl/autotune.py # tests/unit/test_autotune.py

jhinpan · 2026-07-07T06:50:54Z

/rerun-ci

…obustness) Follow-ups from the second multi-model review (round-1 fixes verified): - Cache-hit fallback catches only config-incompatibility errors (ValueError/ TypeError/KeyError); genuine compile/launch/runtime errors now propagate instead of being masked. Logs and drops the stale entry from disk too. - _call_do_bench passes `setup` only when the benchmarker explicitly declares it -- a **kwargs catch-all no longer gets setup passed-and-silently-dropped (which would skip restore/reset). - _save_disk_cache is best-effort: a write failure (read-only FS, full disk) is logged, never crashes an otherwise-successful tune. - _iter_gpu_kernel_funcs applies occupancy hints only to entry-point kernels (gpu.kernel attr), skipping non-kernel device helpers. - _source_fingerprint hashes the source file (transitive over module-level helpers/constants), not just the function body. - Tests: **kwargs do_bench folds setup; cache-hit degrades to default on a stale config; f32 get_all_configs sweeps BLOCK_THREADS. Co-authored-by: Cursor <cursoragent@cursor.com>

jhinpan · 2026-07-07T07:25:14Z

/rerun-ci

jhinpan · 2026-07-07T07:27:15Z

/rerun-ci

jhinpan · 2026-07-08T23:41:12Z

/rerun-ci

Two ir.Module.parse(...) calls were left wrapped multi-line (non-canonical under black's 120-col config, likely after a merge-conflict resolution), which failed the "Check Python Code Style" job. Collapse them to a single line via black. No logic change; ruff was already clean. Co-authored-by: Cursor <cursoragent@cursor.com>

jhinpan · 2026-07-09T00:05:31Z

/rerun-ci

zhiding512 · 2026-07-09T03:41:55Z

+    func_op.attributes["passthrough"] = ir.ArrayAttr.get(kept + [entry])
+
+
+def _apply_occupancy_compile_hints(module: ir.Module) -> None:


Seems we have three methods to set occupancy? Passing gpu-module-to-binary, setting attributes ( in moe) ? We only need one finally.

Good catch — agreed there should be exactly one. Plan: consolidate onto the gpu.func attribute mechanism (the only one the AMDGPU backend actually honors) and drop the rest.

gpu-module-to-binary opts= (--amdgpu-waves-per-eu / --amdgpu-num-vgpr in backends/rocm.py) is the dead path — the backend silently ignores these (as measured earlier in this thread), yet the autotuner was still feeding the same hint into it in parallel with the func-attr lowering. So the hint was actually consumed twice. Removing it so it's consumed in exactly one place. (A unit test in test_external_llvm_codegen.py even asserted these land in opts=; flipping it to assert they don't, to lock the no-op out.)

hand-set value_attrs in MoE/attention and this autotune hint-lowering are really the same mechanism — both end up writing rocdl.waves_per_eu / amdgpu-num-vgpr onto the kernel gpu.func. I'll factor the knob→attribute mapping into a single _set_occupancy_attrs() helper so there's one source of truth; _apply_occupancy_compile_hints routes through it (hand-authored kernels can converge onto it later too).

Net after the change: occupancy hints are consumed in one place, and a single helper defines how a knob becomes a gpu.func attribute.

zhiding512 · 2026-07-09T03:49:13Z

_apply_occupancy_compile_hints applies the same waves_per_eu / maxnreg hints to every gpu.kernel entry point in the module. That’s fine for single-kernel launchers like rmsnorm, but it’s a limitation worth documenting: if a @jit launcher ever emits multiple entry kernels that need different occupancy settings, autotune can’t tune them independently — one Config(waves_per_eu=...) would blanket all kernels.

waves_per_eu / maxnreg were consumed twice: once lowered onto the kernel gpu.func as attributes (the mechanism the AMDGPU backend honors) and once as --amdgpu-waves-per-eu / --amdgpu-num-vgpr on gpu-module-to-binary opts=, which the backend silently ignores. The latter is a dead no-op, so the same hint was being consumed on two paths. - rocm backend: stop forwarding occupancy knobs to gpu-module-to-binary opts= so the hint is consumed in exactly one place. - jit_function: extract _set_occupancy_attrs() as the single source of truth for the knob -> gpu.func attribute mapping; _apply_occupancy_ compile_hints() routes through it, and hand-authored value_attrs kernels land on the same attribute. - test: flip test_rocm_external_pipeline_split_matches_full_pipeline to assert occupancy knobs are absent from opts=, locking out the no-op. Addresses review feedback on ROCm#785 (only one occupancy mechanism). Co-authored-by: Cursor <cursoragent@cursor.com>

waves_per_eu / maxnreg were consumed twice: once lowered onto the kernel gpu.func as attributes (the mechanism the AMDGPU backend honors) and once as --amdgpu-waves-per-eu / --amdgpu-num-vgpr on gpu-module-to-binary opts=, which the backend silently ignores. The latter is a dead no-op, so the same hint was being consumed on two paths. - rocm backend: stop forwarding occupancy knobs to gpu-module-to-binary opts= so the hint is consumed in exactly one place. - jit_function: extract _set_occupancy_attrs() as the single source of truth for the knob -> gpu.func attribute mapping; _apply_occupancy_ compile_hints() routes through it, and hand-authored value_attrs kernels land on the same attribute. - test: flip test_rocm_external_pipeline_split_matches_full_pipeline to assert occupancy knobs are absent from opts=, locking out the no-op. Addresses review feedback on ROCm#785 (only one occupancy mechanism).

Copilot AI review requested due to automatic review settings July 1, 2026 07:05

Copilot started reviewing on behalf of jhinpan July 1, 2026 07:06 View session

Copilot AI reviewed Jul 1, 2026

View reviewed changes

jhinpan mentioned this pull request Jul 1, 2026

[3/5] autotune: offline config emit + runtime lookup (#770) #786

Open

jhinpan force-pushed the feat/autotune-rmsnorm-adopt branch from eb94a0a to 0aa86e0 Compare July 1, 2026 07:56

jhinpan mentioned this pull request Jul 1, 2026

[4/5] autotune: CI guard + committed-config regression check (#770) #788

Open

Merge branch 'main' into feat/autotune-cache-key-restore

50a95c2

jhinpan force-pushed the feat/autotune-rmsnorm-adopt branch 5 times, most recently from e96dad5 to db07800 Compare July 2, 2026 06:54

jhinpan force-pushed the feat/autotune-rmsnorm-adopt branch from db07800 to 95ea147 Compare July 2, 2026 07:08

jhinpan force-pushed the feat/autotune-rmsnorm-adopt branch from 95ea147 to 1bc2ee7 Compare July 2, 2026 19:15

zhiding512 reviewed Jul 4, 2026

View reviewed changes

Comment thread kernels/rmsnorm_config.py Outdated

jhinpan and others added 2 commits July 6, 2026 23:30

Merge branch 'main' into feat/autotune-rmsnorm-adopt (sync to 0.2.3)

c6610e0

jhinpan force-pushed the feat/autotune-rmsnorm-adopt branch from 1e07aff to a19f7e6 Compare July 6, 2026 23:31

jhinpan and others added 3 commits July 7, 2026 00:07

Merge branch 'main' into feat/autotune-rmsnorm-adopt

2cb96df

Merge branch 'main' into feat/autotune-rmsnorm-adopt (sync after ROCm…

a8b87e4

…#783/ROCm#807) Co-authored-by: Cursor <cursoragent@cursor.com> # Conflicts: # python/flydsl/autotune.py # tests/unit/test_autotune.py

Merge branch 'main' into feat/autotune-rmsnorm-adopt

dd2180f

jhinpan added 2 commits July 7, 2026 12:08

Merge branch 'main' into feat/autotune-rmsnorm-adopt

1735782

Merge branch 'main' into feat/autotune-rmsnorm-adopt

3f92925

jhinpan mentioned this pull request Jul 9, 2026

[Feature] Make autotune a real, adopted tuning path (learn from aiter + existing flydsl autotune) #770

Open

6 tasks

zhiding512 reviewed Jul 9, 2026

View reviewed changes

jhinpan force-pushed the feat/autotune-rmsnorm-adopt branch from 4b253f8 to 2a1384e Compare July 9, 2026 04:38

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[2/5] autotune: two-track config + first real adopter (rmsnorm) (#770)#785

[2/5] autotune: two-track config + first real adopter (rmsnorm) (#770)#785
jhinpan wants to merge 16 commits into
ROCm:mainfrom
jhinpan:feat/autotune-rmsnorm-adopt

jhinpan commented Jul 1, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

jhinpan commented Jul 7, 2026

Uh oh!

jhinpan commented Jul 7, 2026

Uh oh!

jhinpan commented Jul 7, 2026

Uh oh!

jhinpan commented Jul 7, 2026

Uh oh!

jhinpan commented Jul 8, 2026

Uh oh!

jhinpan commented Jul 9, 2026

Uh oh!

zhiding512 Jul 9, 2026

Uh oh!

jhinpan Jul 9, 2026

Uh oh!

zhiding512 commented Jul 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		func_op.attributes["passthrough"] = ir.ArrayAttr.get(kept + [entry])


		def _apply_occupancy_compile_hints(module: ir.Module) -> None:

Uh oh!

Conversation

jhinpan commented Jul 1, 2026

Builder mode

Two-track config

First adopter: rmsnorm

Verification (MI350X / gfx950, M=4096 N=8192 bf16)

Tests

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

jhinpan commented Jul 7, 2026

Uh oh!

jhinpan commented Jul 7, 2026

Uh oh!

jhinpan commented Jul 7, 2026

Uh oh!

jhinpan commented Jul 7, 2026

Uh oh!

jhinpan commented Jul 8, 2026

Uh oh!

jhinpan commented Jul 9, 2026

Uh oh!

zhiding512 Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

jhinpan Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

zhiding512 commented Jul 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants