M6: drop mlx::core::compile wrapping after Phase-1 de-risk - #13
Merged
Conversation
A pure-C++ microbenchmark on MLX 0.25.1 / Apple Silicon GPU shows the
fusion win on a Qwen3-0.6B-shaped transformer block is 1.04x-1.07x
(GPU) and a regression on CPU (0.82x-0.88x) — well below the PLAN's
1.20x gate. A sanity workload (pure 8-op elementwise chain) in the
same harness shows 2.78x GPU / 1.47x CPU, so the measurement is
trustworthy — the limiting factor is that MLX compile doesn't fuse
matmul with surrounding elementwise ops, and transformer inference is
matmul-dominated. Fusion wins are confined to RMSNorm / softmax /
SwiGLU neighbourhoods, a small fraction of block runtime.
The BEAM-integrated compile path could not exceed this C++ ceiling,
so Phase 2 (approach review) and Phase 3 (implementation) were not
built.
Retained as permanent fixtures so the decision can be re-measured
against future MLX releases without rebuilding the harness:
- bench/native/compile_microbench.cpp — transformer block + sanity
elementwise chain, both timed compiled vs uncompiled.
- lib/mix/tasks/bench.native.ex — `mix bench.native` invokes the
new `bench-native` Makefile target with the same env elixir_make
sets, reusing the project's pinned MLX without a second fetch.
Supports --seq / --warmup / --iters via `mix bench.native -- ...`.
- Makefile: `bench-native` target, produces
$(BUILD_DIR)/compile_microbench and runs it.
- bench/compile_microbench.md — results table + reproduction.
PLAN.md: M6 section rewritten to record the drop with the measurement;
core design decision #1 and the M5 footnote updated to match.
Emily.Compiler moduledoc note reframed.
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.
Summary
mix bench.native,bench/native/compile_microbench.cpp,bench/compile_microbench.md) as a permanent fixture so the decision can be re-measured against future MLX releases without rebuilding the harness.Emily.Compiler's moduledoc to record the drop and reasoning.Why 1.04× and not more
Transformer inference is matmul-dominated, and MLX's
mx::compiledoes not fuse matmul kernels with surrounding elementwise ops. Fusion is confined to RMSNorm chains, the softmax neighbourhood, and SwiGLU's silu×up — all small fractions of block runtime.A sanity workload (pure 8-op elementwise chain) in the same harness shows 2.78× GPU / 1.47× CPU, confirming the measurement is sound and the harness wiring is correct.
Test plan
mix precommit— compile --warnings-as-errors, format, credo --strict, test. Clean (0 credo issues, 183 tests + 56 properties pass, 0 failures).mix bench.native— produces the full results table (GPU + CPU, transformer + sanity). Committed tobench/compile_microbench.md.mix bench.native -- --seq 512 --iters 500— longer-sequence variant confirms result is not shape-dependent.Artefacts retained for future re-measurement
bench/native/compile_microbench.cpplib/mix/tasks/bench.native.exbench-nativetarget inMakefilebench/compile_microbench.md