Skip to content

Optimize tp fp8 moe for small average router per expert - #4751

Merged
lvhan028 merged 5 commits into
InternLM:mainfrom
grimoire:optimize-fp8moe-smallm
Jul 16, 2026
Merged

Optimize tp fp8 moe for small average router per expert#4751
lvhan028 merged 5 commits into
InternLM:mainfrom
grimoire:optimize-fp8moe-smallm

Conversation

@grimoire

Copy link
Copy Markdown
Collaborator

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily receiving feedbacks. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Please describe the motivation of this PR and the goal you want to achieve through this PR.

Modification

Please briefly describe what modification is made in this PR.

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repositories?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

  1. Pre-commit or other linting tools are used to fix the potential lint issues.
  2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
  3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
  4. The documentation has been modified accordingly, like docstring or example tutorials.

@grimoire
grimoire marked this pull request as ready for review July 15, 2026 07:45
Copilot AI review requested due to automatic review settings July 15, 2026 07:45

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

Pull request overview

This PR adds a “compact routed-block” scheduling path for blocked-FP8 MoE (targeting cases with small average routed tokens per expert) and makes the origin blocked-FP8 launch configuration depend on average routes per expert, aiming to reduce wasted CTAs and improve efficiency.

Changes:

  • Added compact routed-block metadata generation (_get_sorted_idx_blocks) and a new Triton kernel/launcher for compact down-projection.
  • Switched blocked-FP8 MoE to select gate/down configs based on average routes per expert and to conditionally use the compact down path.
  • Added unit tests for config selection, CTA estimation, and compact-down gating logic.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File Description
tests/pytorch/kernel/test_fused_moe.py Adds unit tests for blocked-FP8 MoE config selection and compact-down gating/CTA estimates.
lmdeploy/pytorch/kernels/cuda/fused_moe.py Adds _get_sorted_idx_blocks and Triton kernel to build compact block metadata from routing histograms.
lmdeploy/pytorch/kernels/cuda/blocked_fp8_fused_moe.py Adds compact down-projection kernel/launcher, config selection helpers, and integrates compact down path into fused_moe_blocked_fp8.

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

Comment on lines +560 to +563
if w1.size(0) != w2.size(0):
return False
if w1.size(0) != num_experts:
return False
Comment on lines +622 to +625
gate_moe_cfg, down_moe_cfg = _origin_blocked_fp8_moe_configs(M, topk_ids.numel(), num_experts, E)
use_compact_down = _should_use_compact_blocked_fp8_moe_down(input, input_scale, w1, w1_scale, w2, w2_scale,
topk_ids, num_experts)
if use_compact_down:
return 128


def _origin_blocked_fp8_moe_configs(num_tokens: int, num_routes: int, num_experts: int, local_experts: int):

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

local_experts is not used.

@lvhan028
lvhan028 merged commit 95e3c5a into InternLM:main Jul 16, 2026
3 of 4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants