Skip to content

docs(deepseek_v41): measure disk startup with the merged build - #139

Merged
iamwavecut merged 1 commit into
masterfrom
dev/deepseek-v41-final-numbers
Oct 25, 2026
Merged

iamwavecut merged 1 commit into
masterfrom
dev/deepseek-v41-final-numbers

Conversation

@iamwavecut

Copy link
Copy Markdown
Owner

The merged build (parallel pinned reads, built-in routing profile) was ready in 13 s with disk experts instead of the 20 s the table showed, and decoded the short answer at 8.7 tok/s with the same tokens; host and disk experts score the WikiText slice to the same perplexity (3.067). Documentation only; measured on the RTX 3090 host of #138 before it was released.

@iamwavecut
iamwavecut merged commit b3402a6 into master Oct 25, 2026
1 check failed
@iamwavecut
iamwavecut deleted the dev/deepseek-v41-final-numbers branch October 25, 2026 13:46
13thblade added a commit to 13thblade/ninfer-all-nvfp4 that referenced this pull request Oct 25, 2026
Upstream added five commits on top of 63428ac:

- 6826ac0 perf(artifact): pin host memory with parallel direct reads
- d3b2475 feat(deepseek_v41): serve DeepSeek-V4.1-Flash-1017 from its GSQ-RCO GGUF
- 42376e9 Merge pull request iamwavecut#138 from iamwavecut/dev/deepseek-flash-ninfer-conversion-bce326
- 63b2b23 docs(deepseek_v41): measure disk startup with the merged build
- b3402a6 Merge pull request iamwavecut#139 from iamwavecut/dev/deepseek-v41-final-numbers

88 files changed upstream, 15866 insertions and 445 deletions. Three of them
overlap the two nvfp4 commits' file sets (apps/cli/options.cpp,
src/serve/generation_service.cpp, src/serve/serve_options.cpp) and all three
merged cleanly: upstream only rewraps help text to name DeepSeek-V4.1 beside
Flint.

README.md conflicted in five hunks, every one inside the section that declares
itself "This is the upstream original documentation." Resolution: the fork's own
head (the Chinese section, the English section and the marker, 114 lines) stays
byte-identical, and everything from that marker on is now upstream's README at
b3402a6 verbatim. The mirror therefore tracks upstream again; before this merge
it still carried the pre-Flint wording.

The two PR branches were rebased onto b3402a6 separately:

- pr/nvfp4-w4a16 -> 091cc390, patch-id 8e145ca4 identical to the pre-sync
  commit, 73 files and 1356/257 lines unchanged.
- pr/nvfp4-w4a8-int -> 05c47ff8. ffn.h, ffn.cpp and text.cpp conflicted because
  upstream's tensor-parallel FFN share and this commit's deferred input norm
  both append a trailing parameter to ffn(). The three files are taken byte for
  byte from master, where that collision was already resolved and the two
  objects compiled clean; no new resolution was invented here. 80 of the 96
  files in the commit are byte-identical to the pre-sync tip, and of the other
  16, 13 keep the commit's added lines exactly -- they differ only in upstream
  context. The net diffstat move is +6/-1.

Not covered by this commit: the build, the test suite and any numerics.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant