Skip to content

Skip unnecessary simdgroup computations for quantised MOE matmuls on NAX - #4352

Merged
zcbenz merged 1 commit into
ml-explore:mainfrom
RohanGautam:nax_skip
Aug 20, 2026
Merged

Skip unnecessary simdgroup computations for quantised MOE matmuls on NAX#4352
zcbenz merged 1 commit into
ml-explore:mainfrom
RohanGautam:nax_skip

Conversation

@RohanGautam

@RohanGautam RohanGautam commented Aug 18, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Proposed changes

The kernel used for MOE quantised matmuls (gather_qmm_rhs_nax) does multiple full tile-level matmuls with the corresponding expert matrices when there are more than one expert assigned to a tile, and discards the rows that are not used by that expert. This PR skips the computation for the simdgroups whose computation is going to be discarded anyway at that step.

Benchmarks

The benchmarks were run on M5 Pro (24GB) and M5 max (128GB) using mlx_lm.benchmark . Qwen and OLMoE for affine quantisation (varying number of experts) and gpt-oss for mxfp4 effects.

model machine prompt prefill before prefill after Δ prefill
Qwen3-30B-A3B-4bit M5 Pro 128 857.7 913.9 +6.55%
512 1,896.5 1,977.1 +4.25%
2048 2,766.6 2,909.6 +5.17%
M5 Max 128 1,381.5 1,473.8 +6.68%
512 3,272.8 3,423.8 +4.61%
2048 5,125.7 5,294.2 +3.29%
gpt-oss-20b-MXFP4-Q8 M5 Pro 128 1,139.4 1,203.0 +5.58%
512 2,256.1 2,307.5 +2.28%
2048 3,006.1 3,100.1 +3.13%
M5 Max 128 1,762.0 1,858.5 +5.48%
512 4,131.3 4,223.4 +2.23%
2048 5,308.1 5,502.0 +3.65%
OLMoE-1B-7B-0125-Instruct-4bit M5 Pro 128 2,464.5 2,636.3 +6.97%
512 5,827.1 6,026.4 +3.42%
2048 8,132.6 8,410.3 +3.41%
M5 Max 128 3,936.7 4,193.3 +6.52%
512 9,659.1 10,008.6 +3.62%
2048 13,827.5 14,404.6 +4.17%

Decode speeds are unchanged, as expected.

Checklist

Put an x in the boxes that apply.

  • I have read the CONTRIBUTING document
  • I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
  • I have added tests that prove my fix is effective or that my feature works
  • I have updated the necessary documentation (if needed)

@RohanGautam
RohanGautam marked this pull request as ready for review August 19, 2026 02:05
jagrit06
jagrit06 approved these changes Aug 19, 2026
View reviewed changes

@jagrit06 jagrit06 left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hey! Good catch and work - I have verified it on my M5 machine and it lgtm!

@zcbenz
zcbenz merged commit c7ff35d into ml-explore:main Aug 20, 2026
126 of 140 checks passed
davidtai added a commit to Layr-Labs/mlx that referenced this pull request Aug 25, 2026
@davidtai @JasonHonKL
@AKnassa @zcbenz @Adityaj0 @codeAnqiang-ma @RohanGautam @ayaangazali @erwinzhang7 @katlun-lgtm @robertomeroni @FU-max-boop @HaoXuAI @FeliGame Feli @eyupcanakman @yentur @aleroot @shoutoutuoadi325 @dhiltgen @reckylurker @hojin12312 @x14ngch3n @dudududukim @kapellirohith @devteamaegis @aaishwarymishra @nastya236 @wyanzhao @XXXXRT666 @j-bowhay @vraj00222 @Gusanidas @dwijenpatel @ternaus @deBrian07 Daniel Hiltgen @strayberry katlun-lgtm @anupsv @Gajesh2007 David Tai
chore: upgrade core to MLX 0.32.2 (#10)
734241b 
* Return tuple in meshgrid (ml-explore#4229)

* Add endpoint parameter to linspace (ml-explore#4184)

Co-authored-by: Cheng <git@zcbenz.com>

* Fix vmap of partition/argpartition dropping the kth argument (ml-explore#4116)

* Fix nan_to_num replacing inf with 0 for float16 and bfloat16 (ml-explore#4222)

Co-authored-by: codeAnqiang-ma <273298913+codeAnqiang-ma@users.noreply.github.com>
Co-authored-by: Cheng <git@zcbenz.com>

* Fix einsum not broadcasting batch dimensions in batched tensordot (ml-explore#4125)

Co-authored-by: Cheng <git@zcbenz.com>

* Dequantize in float32 (ml-explore#4241)

* chore: Reject complex in erf and erfinv (ml-explore#4243)

* Fix cpu compilation failure of abs with uint (ml-explore#4240)

Co-authored-by: Cheng <git@zcbenz.com>

* Fix quantize matrix multiplication floor issue (ml-explore#4251)

* Only use MPI backend for world size > 1 (ml-explore#4210)

* chore: Reject complex in expm1, sigmoid and arctan2 (ml-explore#4257)

* Decompose small kernel-depth 3D convs into 2D convs (ml-explore#3785)

Co-authored-by: katlun-lgtm <264247399+katlun-lgtm@users.noreply.github.com>
Co-authored-by: Cheng <git@zcbenz.com>

* Fix Metal sort of a view with a negative stride (ml-explore#4252)

* Mirror the depth axis in the decomposed 3D conv when flipped (ml-explore#4277)

* Fix Metal row reductions on negative-stride views (ml-explore#4267)

Co-authored-by: Fu Xiaonan <214359569+FU-max-boop@users.noreply.github.com>

* [CUDA] Fix custom kernel cache collision for same name, different source (ml-explore#4273)

Co-authored-by: Cheng <git@zcbenz.com>

* Fix ops rejecting integers larger than INT32_MAX (ml-explore#4255)

Co-authored-by: Feli <feli@hnu.edu.cn>
Co-authored-by: Cheng <git@zcbenz.com>

* Fix var/std for complex numbers (ml-explore#4260)

* Fix int32 overflow in conv padded input and pad shapes (ml-explore#4258)

Co-authored-by: Cheng <git@zcbenz.com>

* chore: Reject complex in remainder (ml-explore#4270)

* chore: Compare the macOS SDK version as a version when gating JACCL (ml-explore#4286)

* Clamp ring socket transfers so a payload of 2 GiB or more can be sent (ml-explore#4281)

Co-authored-by: Cheng <git@zcbenz.com>

* chore: Use normalize_axis_index in split/unstack/partition/topk (ml-explore#4288)

* Remove grouped output in CI (ml-explore#4195)

* [CUDA] Fix finding cuda 13 headers in JIT compilation (ml-explore#3995)

* Refactor wheel building script (ml-explore#3818)

* Make mx.compile cache erasing thread safe (ml-explore#4248)

Co-authored-by: yentur <mr.yentur@gmail.com>

* Add builds for free-threaded python (ml-explore#3812)

* Fix int32 overflow in concatenate/repeat/kron (ml-explore#4303)

* python: Widen list elements that do not fit in int32 to int64 (ml-explore#4305)

* Propagate CPU errors to events (ml-explore#3742)

Co-authored-by: Alessio Pollero <alessio.pollero@gmail.com>

* Fix mx.arange dtype inference overflow regression (ml-explore#4324)

* Add workflow to update pull request limit bypass list (ml-explore#4320)

* Support head dimension 72 in Metal full attention (ml-explore#4330)

* Patch bump to 0.32.2 (ml-explore#4333)

* Preserve subnormal float values when casting to bool (ml-explore#4224)

* python: Support assigning through a bare Ellipsis index (ml-explore#4314)

* Fix divmod truncating the quotient for floats (ml-explore#4108)

Co-authored-by: Cheng <git@zcbenz.com>

* Add force_fused option to scaled_dot_product_attention (ml-explore#4185)

* chore: Reject negative eps in the normalization layers (ml-explore#4312)

* Bound GGUF metadata string/array values against the file mapping (ml-explore#4212)

Co-authored-by: x14ngch3n <x14ngch3n@users.noreply.github.com>
Co-authored-by: Cheng <git@zcbenz.com>

* Read each K/V byte once in gqa-8 decode attention (ml-explore#4077)

* Fix fft vmap and jvp for transforms over a subset of axes (ml-explore#4138)

* Fix median dropping NaN (ml-explore#4146)

* Fix the CPU scan over a size one axis with a padded stride (ml-explore#4139)

Co-authored-by: Cheng <git@zcbenz.com>

* chore: Validate the optimizer betas at construction (ml-explore#4310)

Co-authored-by: Cheng <git@zcbenz.com>

* `RMSNormVJP` backward writes a full `{n_rows, D}` `gw_temp` intermediate (ml-explore#4293)

* [Bug]: add default none value to axis parameter of the take_along_axis (ml-explore#4357)

Co-authored-by: Anastasiia Filippova <a_filippova@apple.com>

* Add a fused full-attention path for head_dim 256 on NAX devices (ml-explore#3842)

Co-authored-by: Cheng <git@zcbenz.com>

* Update nanobind to 2.15.0 (ml-explore#4337)

* Skip unnecessary simdgroup computations for quantised MOE matmuls on NAX (ml-explore#4352)

* Add AI usage policy (ml-explore#4331)

Co-authored-by: Jake Bowhay <60778417+j-bowhay@users.noreply.github.com>

* Raise cpu stream errors from synchronize (ml-explore#4338)

Co-authored-by: Cheng <git@zcbenz.com>

* chore: Validate eps in Adam at construction (ml-explore#4361)

Co-authored-by: Anastasiia Filippova <a_filippova@apple.com>

* Bound winograd conv2d working set by tiling the batch (ml-explore#4102)

Co-authored-by: Cheng <git@zcbenz.com>

* Use a 32-row block in qmm_t_nax when one block covers all of M (ml-explore#4171)

* chore: Deduplicate fftshift and ifftshift (ml-explore#4318)

* Fix Log and Equal is_equivalent ignoring primitive state (ml-explore#4266)

Co-authored-by: Cheng <git@zcbenz.com>

* Stabilize reduced-precision InstanceNorm (ml-explore#4230)

* chore: Normalize negative axes in sort and argsort (ml-explore#4332)

* Clean up main thread compile cache before python interpreter shuts down (ml-explore#4373)

* chore: Check malformed jaccl hostfile that miss rdma in pairs (ml-explore#4284)

Co-authored-by: Cheng <git@zcbenz.com>

* Round mxfp8 block scales up to avoid saturation (ml-explore#4353)

Co-authored-by: Daniel Hiltgen <daniel.hiltgen@ollama.com>
Co-authored-by: Cheng <git@zcbenz.com>

* Add support for the __array_namespace_info__  (ml-explore#4334)

* Stop a failed CUDA graph commit from poisoning the encoder (ml-explore#4356)

Co-authored-by: Cheng <git@zcbenz.com>

* Fix quantized kernels in JIT build (ml-explore#4372)

Co-authored-by: Cheng <git@zcbenz.com>

* Avoid zero work in stride-2 ConvTranspose3d (ml-explore#4343)

* [CUDA] Ce fused kernel (ml-explore#3947)

* Fix cpu exclusive scan for complex numbers (ml-explore#4272)

Co-authored-by: Cheng <git@zcbenz.com>

* Support Relocatable CUDA DLLs on Windows (ml-explore#4382)

* Use cast_to for fused AsType in compiled Metal kernels (ml-explore#4351)

Co-authored-by: katlun-lgtm <katlun@windyviews.com>
Co-authored-by: Cheng <zcbenz@gmail.com>

* python: Declare DLPackCompatible protocol members as methods (ml-explore#4384)

* Fix quantizing sliced arrays (ml-explore#4381)

* Fix einsum dropping a trailing empty subscript (ml-explore#4299)

Co-authored-by: Cheng <git@zcbenz.com>

* Add script to run python tests (ml-explore#4393)

* Hold GIL in AttachedData destructor (ml-explore#4391)

* Bound Metal buffer COUNT, not just bytes, in MetalAllocator

The Metal allocator throws `[metal::malloc] Resource limit (N) exceeded`
when num_resources_ (the live+cached Metal buffer COUNT) reaches
resource_limit_ (the iogpu.rsrc_limit sysctl, default ~499000). Freed
buffers are recycled into a size-keyed cache whose only trim is by BYTES
(release_cached_buffers takes a bytes-to-free target, max_pool_size_ ~=
physical RAM). Under churn with many distinct buffer shapes (varied prompt
lengths, growing KV caches, multiple co-resident models) the cache fills
with entries never reused at that exact size, so the COUNT climbs to the
limit while byte usage stays modest and the byte trim never fires — the
process crashes mid-inference on a machine with most of its RAM free.

malloc() now also reclaims by count: when num_resources_ crosses a 90%
high-water mark of resource_limit_, it clears the (pure-reuse) buffer
cache so the count drops back to the live working set. Clearing the cache
only costs re-allocation, never correctness, so the count limit becomes
unreachable by any request mix or batching method while the existing byte
limits keep total memory bounded.

Adds get_num_resources()/get_resource_limit() to the public memory API
(metal + no_gpu + cuda backends) so the count and its ceiling are
observable from callers. Adds an MLX_RESOURCE_LIMIT env override that can
only LOWER the ceiling (clamped to the OS limit, strictly validated) to
exercise the trim deterministically and as an operator safety valve.

* perf(mlx): opt-in Gemma 4 expert-QMM tile kernel with parallel descriptor builder (#4)

* perf(mlx): add opt-in Gemma 4 expert-QMM tile kernel with parallel descriptor builder

Adds a distinctly-named expert QMM implementation for the Gemma 4
26B-A4B MoE production shapes, gated by MLX_GATHER_QMM_EXPERT_SLICES:

- qmm_t_expert_impl: BM32 expert tile body (BM16 fallback rows) taking a
  private/by-value row count; the shared qmm_t_impl constant-address ABI
  and all ordinary gathered/batched/dense QMM routes are unchanged.
- build_gemma4_sorted_expert_tiles_bm32: one 128-thread threadgroup
  replaces the reference design's single-GPU-thread serial builder;
  parallel expert-range binary search, Hillis-Steele scan, and strided
  upper-bound descriptor emission.
- Selector runs after the NAX-first route and requires affine BF16
  transposed inputs, 4-bit gs=64 weights, 128 experts, assignment counts
  of exactly 4096/8192/16384, and the exact gate/up or down rank-3
  shapes; every miss keeps the legacy route. NAX engagement is
  non-engagement, never bypassed.
- device.{h,cpp}: one-shot request resolution, nonthrowing dual-symbol
  AOT probe/prewarm, relaxed-atomic diagnostics (requested, aotAvailable,
  naxAvailable, hits, per-class fallbacks).
- gpu_tests: exact-shape arithmetic parity, fallback, and counter
  invariant probes.

Retention standing (2026-08-09 production matrix): opt-in experiment.
Standalone profile dropped (prefill -10.2% vs bracket); paired
weighted-unsort+R1 profile retained-final (prefill +1.8%, TTFT -7.5%,
decode +3.3%, arrival E2E +12.0%). NOTE: this source post-dates the
benchmarked binaries/metallib (post-measurement kernel-body edit);
rebuild and re-verify before any performance claim.

* fix(mlx): fail-safe sortedness check in gemma expert tile builder; counter/atomic hygiene

Review-wave fixes for the R1 expert-QMM path:

- N1 (sortedness trust): build_gemma4_sorted_expert_tiles_bm32 now
  verifies each thread's post-binary-search segment boundary against the
  generalized invariant indices[start - 1] < lid <= indices[start]
  (edge threads check their single neighbor), votes per simdgroup via
  simd_or, folds the votes through threadgroup memory, and on any
  violation retracts count[0] to 0 (tile kernel then early-returns) and
  records the violation in count[1]; the buffer ABI is unchanged
  (count index 1 was previously unused). try_gemma4_expert_qmm allocates
  the second count element, drains the encoder after the builder, and
  re-routes a retracted call to the order-agnostic legacy path instead of
  dispatching the tile kernel (zero count is unambiguous: the selector's
  assignment gate guarantees M is 4096/8192/16384).
- N2 (route-condition duplication): the sorted-RHS gate literal that
  appeared (negated) in the diagnostics record and in the dispatch
  decision is now the shared static constexpr predicate
  takes_sorted_rhs_route, so future tuning of the 16/4 thresholds cannot
  desynchronize counter vs route.
- N3 (per-call bias normalization): gather_qmm_rhs no longer spends
  ensure_row_contiguous on biases before classification reads the raw
  tensor's fields; normalization runs only inside the winning-route
  branch (hit semantics unchanged; the legacy block keeps its own
  normalization point and ordering).
- N4 (armed_ data race): Gemma4ExpertQMMCounters::armed_ is now
  std::atomic<bool> with relaxed loads/stores in armed(), snapshot(),
  snapshot_and_disarm() (read-then-write order preserved) and
  clear_and_arm(); the class remains non-copyable, now enforced.

* fix(mlx): make the R1 sortedness fail-safe sound; proper retract attribution

F1: the per-expert boundary vote was a partial detector -- an inversion
inside a segment used by no other expert's boundary could escape, so
"re-route on any violation" overclaimed. build_gemma4_sorted_expert_tiles_bm32
now also runs a strided adjacent-pair scan: thread lid checks
indices[i-1] <= indices[i] for i = lid+1; i < M; i += 128, covering every
adjacent pair in [1, M) exactly once (1..128 iterations at the reachable
M in {4096,8192,16384}). Adjacent-pair monotonicity is transitive, so a
clean scan is a sound and complete sortedness oracle; it folds into the
same simd_or/threadgroup vote and the same retract (count[0]=0, count[1]=1).
The boundary checks stay as cheap, precise diagnostics.

F2: retracts were write-only in count[1] and surfaced as
fallback_metallib_unavailable -- misattribution in the only observable
surface. A dedicated fallback_sortedness_retracted counter now rides the
GemmA4 route counters and the C diagnostics ABI
(sizeof 80 -> 88, new uint64 at offset 80; existing offsets unchanged).
try_gemma4_expert_qmm returns the route class: count[0]==0 with count[1]==1
records fallback_sortedness_retracted, any other unusable build keeps
fallback_metallib_unavailable, then re-routes to the legacy path as before.

F4: new doctest drives the full armed() -> clear_and_arm() ->
snapshot_and_disarm() cycle and the attempts == hits + fallbacks invariant
including the new class; the route-table and counter-invariant tests now
cover fallback_sortedness_retracted.

Verified: cmake tests 262/262 + 3550 assertions pass; metal -Wall -Wextra
-fno-fast-math compile of kernels/quantized.metal is warning-free.

* perf(metal): E=256 expert-tile route + trust + gpu::eval UAF fix — darkbloom-base mirror (#7)

* perf(metal): instantiate E=256 expert-tile route for Qwen 3.5/3.6 MoE prefill (mirror of Cmlx/mlx 58fab46)

* fix(metal): use-after-free in gpu::eval for primitives that synchronize mid-eval (mirror)

* perf(metal): trust mode skips retract readback (mirror)

* fix(compile): preserve all-cache binding cleanup

---------

Co-authored-by: JasonHonKL <148705846+JasonHonKL@users.noreply.github.com>
Co-authored-by: AK <144495202+AKnassa@users.noreply.github.com>
Co-authored-by: Cheng <git@zcbenz.com>
Co-authored-by: Adityaj0 <93090622+Adityaj0@users.noreply.github.com>
Co-authored-by: anchor <codeanqiang@gmail.com>
Co-authored-by: codeAnqiang-ma <273298913+codeAnqiang-ma@users.noreply.github.com>
Co-authored-by: Rohan Gautam <rohan1gautam@gmail.com>
Co-authored-by: Ayaan Gazali <ayaangazali.work@gmail.com>
Co-authored-by: Erwin Zhang <59893706+erwinzhang7@users.noreply.github.com>
Co-authored-by: katlun-lgtm <katlun@gmail.com>
Co-authored-by: katlun-lgtm <264247399+katlun-lgtm@users.noreply.github.com>
Co-authored-by: robertomeroni <150194833+robertomeroni@users.noreply.github.com>
Co-authored-by: Fu Xiaonan <ht3fudatou@163.com>
Co-authored-by: Fu Xiaonan <214359569+FU-max-boop@users.noreply.github.com>
Co-authored-by: Hao Xu <hxu44@apple.com>
Co-authored-by: Feli <89400571+FeliGame@users.noreply.github.com>
Co-authored-by: Feli <feli@hnu.edu.cn>
Co-authored-by: Eyüp Can Akman <eyupcanakman@gmail.com>
Co-authored-by: Cheng <zcbenz@gmail.com>
Co-authored-by: yentur <mr.yentur@gmail.com>
Co-authored-by: Alessio Pollero <alessio.pollero@gmail.com>
Co-authored-by: Zhiqi Zhang <zhiqizhangg@gmail.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
Co-authored-by: Tanish Jain <recklurker@gmail.com>
Co-authored-by: hojin12312 <hojin12312@gmail.com>
Co-authored-by: Xiang Chen <46052474+x14ngch3n@users.noreply.github.com>
Co-authored-by: x14ngch3n <x14ngch3n@users.noreply.github.com>
Co-authored-by: Duhyeon, Kim <49020301+dudududukim@users.noreply.github.com>
Co-authored-by: rohith <kapellirohith@gmail.com>
Co-authored-by: Ishaan Samantray <devteam.aegis@gmail.com>
Co-authored-by: Aaishwarya Mishra <aaishwarymishra@gmail.com>
Co-authored-by: Anastasiia Filippova <a_filippova@apple.com>
Co-authored-by: Yanzhao Wang <19340816+wyanzhao@users.noreply.github.com>
Co-authored-by: XXXXRT666 <157766680+XXXXRT666@users.noreply.github.com>
Co-authored-by: Jake Bowhay <60778417+j-bowhay@users.noreply.github.com>
Co-authored-by: vraj patel <87225460+vraj00222@users.noreply.github.com>
Co-authored-by: Gusanidas <33495733+Gusanidas@users.noreply.github.com>
Co-authored-by: Dwijen Patel <dwijen@gmail.com>
Co-authored-by: Vladimir Iglovikov <ternaus@users.noreply.github.com>
Co-authored-by: Brian C. <94733710+deBrian07@users.noreply.github.com>
Co-authored-by: Daniel Hiltgen <daniel.hiltgen@ollama.com>
Co-authored-by: YH Yan <strayberry0w0@gmail.com>
Co-authored-by: katlun-lgtm <katlun@windyviews.com>
Co-authored-by: anupsv <6407789+anupsv@users.noreply.github.com>
Co-authored-by: Gajesh Naik <26431906+Gajesh2007@users.noreply.github.com>
Co-authored-by: David Tai <davidtai@Davids-MBP.lan>
This was referenced Aug 25, 2026
Open
Draft
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@jagrit06 jagrit06 jagrit06 approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@RohanGautam @jagrit06 @zcbenz