Repository navigation
docs(musa): Moore Threads support matrix, vendor section and pinned torch_musa 2.7.1 profile - #44
Merged
Merged
Conversation
Adds MUSA to the compatibility tables and a measured vendor section. Measured on MTT S70 (capability (2,1)) with torch 2.7.1 + torch_musa 2.7.1 + MUSA SDK 4.3.0.CC2.1 + driver 3.3.0; this port targets 2.7.1 by default. * README.md / README_CN.md: new MUSA row (`musa`/`moore`, torch_musa 2.7.1); * docs/cn-gpu.md: MUSA row, auto-detection order, `PYMSS_DEVICE_BACKEND` value, and a section covering the 春晓 (Chunxiao) version ceiling, the measured capability gaps (complex scatter/index, complex pad, fp16 conv, flash SDPA, half muFFT), the 11-family benchmark at 10 s audio, and the traps we hit; * the fused RoPE kernel turned out to be worth using (7.6x, mel_band 3.31 -> 3.09 s, bs_roformer 2.88 -> 2.14 s), so the section records how to feed it (per-element angle table with `rotary_interleaved=True`) and why its output layout is equivalent to the reference path; * the other vendor-only operators are recorded as measured-and-rejected, with the numbers that justify it; * the section deliberately does not document how to build torch_musa -- installing the vendor torch is the deployer's job -- and does not compare against older torch_musa releases, since 2.7.1 is the baseline.
The profile contract still asserted the torch_musa 2.5.0 workarounds. On 2.7.1 the native paths work (rfft/irfft 1.9e-07, complex-by-real 0, complex cat/fill, `_thnn_fused_gru_cell`, `_native_multi_head_attention`), so those flags flip to their native defaults and only the three real gaps are left: no complex `scatter_add_`/`index_add_`, no fp16 FFT, and fp16 autocast aborting on real models (MUDNN's fused fp16 conv, bfloat16 datacast). Measured on MTT S70 + torch 2.7.1 + torch_musa 2.7.1 + MUSA SDK 4.3.0.
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
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.
Follow-up to #40, which was squash-merged: the 66 commits it carried became a single commit on
main, so this branch had to be rebased onto the newmainbefore it could be proposed again.cn-gpu-pris nowmain+ 2 commits, nothing else.What
Adds Moore Threads (MUSA) to the compatibility matrix and pins the measured
torch_musa 2.7.1profile.Measured on MTT S70 (capability
(2,1)) with torch 2.7.1 + torch_musa 2.7.1 + MUSA SDK 4.3.0.CC2.1 + driver 3.3.0; this port targets 2.7.1 by default.Changes
README.md/README_CN.md: new MUSA row (musa/moore, torch_musa 2.7.1).docs/cn-gpu.md(+80): MUSA row, auto-detection order,PYMSS_DEVICE_BACKENDvalue, and a measured vendor section coveringrotary_interleaved=True) and why its output layout is equivalent to the reference path,It deliberately does not document how to build torch_musa (installing the vendor torch is the deployer's job) and does not compare against older
torch_musareleases, since 2.7.1 is the baseline.test/test_backends_musa.py: the profile assertions pinned to the measurements above.Note on the branch
upstream/main(7d599ea) has the same tree as the old PR head (b61049c), so the rebase was conflict-free andgit diff <old head> <new head>is empty — the content is byte-identical to what was reviewed before, minus the commits that are already inmain.