Repository navigation
[sampler cold start] Load DeepSeek-V4.1-Flash sampler weights in parallel - #148
Merged
Merged
Conversation
This was referenced Oct 28, 2026
micahtyong
changed the title
Base automatically changed from
micah/sampler-perf/S2-running-requests-rev2
to
micah/sampler-perf/S1-kernel-cache
October 28, 2026 17:32
…volume (#146) Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
The FP8/FP4 release is 476 GiB in 48 shards and still loads through mmap, so eight loader threads halve the shard loop without raising host memory. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
micahtyong
force-pushed
the
micah/sampler-perf/S3-multithread-load-rev2
branch
from
October 28, 2026 23:45
47cba94 to
bec9001
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
DeepSeek-V4.1-Flash samplers now read their checkpoint with eight threads instead of one.
The recipe used one thread because the earlier BF16 export (#75) only fit in host memory that way. The FP8/FP4 release the recipe loads today is read through the page cache, which the eight GPU ranks of a sampler share. Eight threads, SGLang's default, finish the read in about half the time without using more host memory.
E2E results
Setup:
deepseek_v41_flash_1017_lora_8kwith one 8×H200 sampler replica, time to read the 48 checkpoint shards (476 GiB) during sampler start, from the same GRPO runs as #146. The other PRs in the stack don't change how shards are read. Before / After🤖 Generated with Claude Code