Repository navigation
[CI] Run test_unified_radix_cache_kl_dcp on cutedsl_mla with bf16 KV cache - #39553
Merged
Merged
Conversation
kpham-sgl
added
run-ci
Collaborator
Author
|
/rerun-test test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py |
Contributor
|
…cache The test has been failing ~50% of scheduled extra-b B200 runs since 09/11 with avg_kl_div 0.010-0.013 vs the 0.01 threshold. sgl-project#35770 routed Kimi Linear prefill through the tokenspeed_mla FP8 prepare_prefill_qkv hook, which doubled the KL from 0.003-0.006. Switch to cutedsl_mla on the default bf16 KV cache like every other KL test; threshold unchanged.
kpham-sgl
force-pushed
the
kpham/kl-dcp-cutedsl
branch
from
September 15, 2026 06:45
12bce74 to
1298ff3
Compare
Collaborator
Author
|
/rerun-test test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py |
Contributor
|
Results for 🚀 |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
test_unified_radix_cache_kl_dcp.pyhas failed ~50% of scheduledextra-b-test-4-gpu-b200runs since 09/11 (avg_kl_div0.010-0.013 vs threshold 0.01, both retry attempts).Bisected to #35770:
resolve_attn_backend()now unwrapsHybridLinearAttnBackend.full_attn_backend, so Kimi Linear prefill hitsTokenspeedMLABackend.prepare_prefill_qkvand quantizes Q/K/V to FP8 before the prefill kernel (#38152 made the hook NoPE-safe after it first crashed on 09/05). Scheduled runs measured 0.003-0.006 through 09/04 and 0.007-0.013 since. Not runner-specific; no package versions changed.Modifications
Switch the test to
cutedsl_mlaon the default bf16 KV cache (drop--kv-cache-dtype fp8_e4m3), like every other registered KL test.cutedsl_mlais the DCP-native MLA decode kernel on Blackwell and already serves Kimi Linear DCP intest_kimi_linear_unified_memory_dcp_blackwell.py. Threshold unchanged at 0.01.Test
10 consecutive runs on 4x B200, main @ 2fca6d6: 10/10 OK, ~4.5 min each.
GSM8K 0.930-0.945 (threshold 0.85). Same-config tokenspeed_mla runs on scheduled CI since 09/06: 0.007-0.013.