[FlyDSL][MOE] Enable a8w8 blockscale moe splitk in flydsl #3280
Open
lalala-sh wants to merge 1 commit into
Open
[FlyDSL][MOE] Enable a8w8 blockscale moe splitk in flydsl #3280lalala-sh wants to merge 1 commit into
lalala-sh wants to merge 1 commit into
Conversation
Contributor
🏷️ CI GuideRuns automatically on every PR:
Extended tests (opt-in via labels):
|
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
| if out_dtype not in ("f16", "bf16"): | ||
| raise ValueError(f"out_dtype must be 'f16' or 'bf16', got {out_dtype!r}") | ||
| # NOTE: don't materialize MLIR types outside an active MLIR Context. | ||
| out_mlir = lambda: (lambda ty: ty() if callable(ty) else ty)(T.f16 if out_dtype == "f16" else T.bf16) |
Contributor
There was a problem hiding this comment.
Do not assign a lambda expression, use a def
Suggested change
| out_mlir = lambda: (lambda ty: ty() if callable(ty) else ty)(T.f16 if out_dtype == "f16" else T.bf16) | |
| def out_mlir(): | |
| return (lambda ty: ty() if callable(ty) else ty)(T.f16 if out_dtype == "f16" else T.bf16) |
| elem_type_tag = "bf16" | ||
| else: | ||
| raise ValueError(f"Unsupported dtype: {dtype_str}") | ||
| compute_type = lambda: T.f32 |
Contributor
| else: | ||
| raise ValueError(f"Unsupported dtype: {dtype_str}") | ||
| compute_type = lambda: T.f32 | ||
| i8_type = lambda: T.i8 |
Contributor
Comment on lines
+2911
to
+2912
| a_dtype_str = "fp8" | ||
|
|
Contributor
Collaborator
|
@lalala-sh ck ? or flydsl? |
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
Enable the FlyDSL backend for a8w8 FP8 blockscale (
per_1x128/per_128x128) MoE infused_moe, and provide the FlyDSL stage1/2 blockscale kernels + tuner integration + tuned configs for the four dsv3 v3 shapes ((model_dim=7168, inter_dim={256,512})×(E,topk)={(256,8), (257,9)}).For small-token decode (M ≤ 4) the FlyDSL 2-stage path now consistently beats the ASM 1-stage blockscale kernel on gfx950 (e.g. M=1: 23.6 us vs 26.6 us, ~13% faster); for medium / large M the tuner still picks the ASM 1-stage where it is faster, so the change is strictly opt-in via the tuned CSV.
Technical Details
Test Plan
Test Result
Submission Checklist