test(pp): add qwen3_5_moe PP and EP parity tests [3/4] - #3649
Open
athitten wants to merge 1 commit into
Open
Conversation
Contributor
Author
|
/ok to test dfa6927 |
athitten
force-pushed
the
athitten/test/parallelism-qwen3-5-moe
branch
from
August 27, 2026 17:03
dfa6927 to
36e9b10
Compare
athitten
changed the title
Qwen3.5-MoE recipes ship pp_size 2-8 and ep_size 4-32, and nothing checked that splitting the pipeline or sharding the experts leaves the numbers unchanged. Adds a small proxy config plus two tests that run it twice and compare the loss and gradient-norm curves: PP2 against a single-rank baseline, and EP2 against a 2-rank data-parallel run. The shipped recipes are VLM, but no Qwen3.5 processor is staged in CI, so the proxy uses the text-only Qwen3_5MoeForCausalLM and generates its own tokens. Nothing has to be staged to run these. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Abhishree <abhishreetm@gmail.com>
athitten
force-pushed
the
athitten/test/parallelism-qwen3-5-moe
branch
from
August 27, 2026 17:37
36e9b10 to
e5042f1
Compare
Contributor
Author
|
/ok to test e5042f1 |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
Two parity tests for Qwen3.5-MoE, plus the small config they run.
L2_Parallelism_Qwen3_5MoE_PP2_Parity.shL2_Parallelism_Qwen3_5MoE_EP2_Parity.shqwen3_5_moe_proxy.yamlqwen3_5_35b_neat_packing.yaml(6 layers, 8 experts).Why
Qwen3.5-MoE recipes ship
pp_size2–8 andep_size4–32. Nothing checked that splitting the pipeline or sharding the experts leaves the numbers unchanged. CP is already covered by the tests undertests/functional_tests/context_parallel/.Nothing to stage
The shipped recipes are VLM, and there is no Qwen3.5 processor in
TEST_DATA_DIR. So the proxy uses the text-onlyQwen3_5MoeForCausalLMand generates its own tokens — no processor, no dataset, no tokenizer. Expert routing, expert sharding and the pipeline split all live in the text part, so they are covered. The vision tower is not.Measured
EP is tighter because both its runs are 2-rank and wrapped the same way, so only expert sharding differs. The PP baseline runs on 1 rank without FSDP, and that difference — not pipeline parallelism — is what widens its numbers.
🤖 Generated with Claude Code