Repository navigation
Conversation
…ob_read predicate Assisted-by: Claude Opus 5.5 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Td2gieja9Da2rPTN17Z846
Assisted-by: Claude Opus 5.5 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Td2gieja9Da2rPTN17Z846
An isolated read matches with max_len 0 (no host-tier lookup) and donates nothing. The prefill-commit branch also skips the snapshot insert for a non-donating read; the scheduler only reaches it for requests that keep decoding, never a read. Assisted-by: Claude Opus 5.5 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Td2gieja9Da2rPTN17Z846
gdevenyi
force-pushed
the
5080/48-logprob-read-isolate
branch
from
October 21, 2026 02:05
adbf2b3 to
cc13fec
Compare
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Makes a logprob read's scores independent of other traffic, behind
--logprob-read-isolate. A read is a request withlogprobsandmax_tokens == 1, the shape AnyJev scoring sends. The flag is off by default; production turns it on. Stacked on #95.Commits
fc305238: theSamplingParams.is_logprob_readpredicate and the flag.fed01568: a read is prefilled in a pass of its own. It is never co-batched, so cuBLAS kernel choice by total M, the per-tensor FP8 amax and the MoE configuration all depend only on its own prompt. A chunked read's continuation is also alone.cc13fecc: reads neither resume from the prefix cache nor donate to it (max_len=0, no host-tier lookup, donatenone). Also guards the chunk-commit branch undernone. That gap is not reachable from the scheduler today, but it is now closed inCacheManager.Why
Batched vs single prefill differs by 0.4-1.7 nats on top-5 logprobs. The cause is GEMM kernel selection by total row count, plus per-tensor FP8 scales shared across the batch (fork #94). A prefix-cache resume also changes numerics compared with a cold prefill. So the same AnyJev question got different scores depending on what else was running and on what ran before it.
GPU check (
integ7/invariance/read_checks.py,batch_debug.py,anyjev_check.py)Cost: short parallel reads are about 1.8x slower, and rotation scoring is about 2x slower because there is no prefix reuse. Chat and side calls run exactly the same code path as before.
CPU tests:
tests/scheduler/test_scheduler_chunked_prefill.pyandtest_hybrid_cache_manager.py. They fail before (6 and 4 failures) and pass after.🤖 Generated with Claude Code
https://claude.ai/code/session_01Td2gieja9Da2rPTN17Z846