Tier 5: Wave C — MoE-decode perf + response_format json + Sharp v22.3.2 template (terse kwarg) - #6

Merged

Gevil merged 139 commits into

Aug 24, 2026

Commits on Aug 19, 2026

fix(serve): name the exception that terminates the process
Sociopacific
committed
fff06a9
feat(serve): report prefix-cache hits in the chat-completions usage
Sociopacific
committed
7503d8a
fix(serving): accept content-part arrays for tool messages
jtf9808
committed
7024e8d
Merge branch 'Neroued:master' into fix/tool-message-content-parts
jtf9808
committed
3cecdf4
fix(serving): generalize parse_content_parts text_only into allowed_types allowlist
jtf9808
committed
bdc0527
fix(serving): export parse_content_parts and unit-test the allowed_types allowlist
jtf9808
committed
f197be4

Commits on Aug 20, 2026

feat(serve): bound each image with a Vision-token budget

Sociopacific
and
claude
committed
eb413c7
perf(ops): fuse sigmoid gate into attention reduce epilogue

ARVLab
authored and
MichaelDementii
committed
ed505eb
perf(ops): fuse q/k rmsnorm into one decode kernel

ARVLab
authored and
MichaelDementii
committed
d3278b7
perf(ops): fold the moe router selection into the last D1 block
MichaelDementii
committed
a2de4e3

Commits on Aug 21, 2026

perf(ops): prefetch next projection weights from moe d4 tail

ARVLab
authored and
MichaelDementii
committed
4514563
perf(ops): prefetch shared expert down weights behind moe router

ARVLab
authored and
MichaelDementii
committed
5750f15
build(win32): add native Windows build support

natpate
authored and
cometkim
committed
883fe85
fix(platform): make artifact reader, load progress, and console log portable to Windows

natpate
authored and
cometkim
committed
92245e2
build(win32): route library dependencies through the platform targets

natpate
authored and
cometkim
committed
8c1b27d
fix(platform): portable media acquisition and request log test

natpate
authored and
cometkim
committed
8df6f62
docs(win32): add Windows build guide and document the Windows path

natpate
authored and
cometkim
committed
1ab6198
docs(readme): document Windows 11 as a supported build platform

natpate
authored and
cometkim
committed
0cd24be
docs: point build instructions at the ninfer-windows repository

natpate
authored and
cometkim
committed
28f3dfe
docs(win32): point build guide at the ninfer-windows repository

natpate
authored and
cometkim
committed
ad8c471
build(win32): permit over-aligned CUDA parameters for the TMA kernels

natpate
authored and
cometkim
committed
5ed5da2
Revert "build(win32): permit over-aligned CUDA parameters for the TMA kernels"

natpate
authored and
cometkim
committed
3491cb4
fix(ops): pass NVFP4 TMA descriptors by device pointer on MSVC

natpate
authored and
cometkim
committed
7757d81
fix(ops): use pair_row in the SwiGLU TMA epilogue

natpate
authored and
cometkim
committed
57aae89
fix(targets): remove out-of-line qwen3_6 plan move specializations

natpate
authored and
cometkim
committed
eab1919
fix(targets): default qwen3_6 plan moves in-class

natpate
authored and
cometkim
committed
97d7d37
fix(targets): keep qwen3_6 plan moves declared, not defaulted, in runtime.h

natpate
authored and
cometkim
committed
841875f
fix(targets): define qwen3_6 plan moves with explicit bodies

natpate
authored and
cometkim
committed
f004960
serve: accept stock llama.cpp WebUI dialect (API-compatible with tools/ui)

natpate
authored and
cometkim
committed
ca0763b
serve: add --webui / --webui-dir to serve the stock llama.cpp WebUI in-process

natpate
authored and
cometkim
committed
b2aec4d
fix(webui): strip scheme from host in split_url

natpate
authored and
cometkim
committed
f9f1f3d
docs(readme): move Upstream section up, list fork additions, add prebuilt release instructions

natpate
authored and
cometkim
committed
8e4111e
docs: make the prebuilt-release section version-agnostic

natpate
authored and
cometkim
committed
423bc27
docs: document per-model context ceilings in the release section

natpate
authored and
cometkim
committed
3cd1120
feat(serve): report served context ceiling as meta.n_ctx on /v1/models objects

natpate
authored and
cometkim
committed
3ab8e8e
fix(serve): brace ordering in models-list meta payload

natpate
authored and
cometkim
committed
23016bc
docs(readme): list context window reporting in the upstream section

natpate
authored and
cometkim
committed
12a577d
Update README.md

natpate
authored and
cometkim
committed
a9ead9f
Update README.md

natpate
authored and
cometkim
committed
82b50b8
Update README.md

natpate
authored and
cometkim
committed
657b42d
fix(build): guard winhttp webui updater to windows
cometkim
committed
7f90fc1
feat(target): register qwen3.8-27b nvfp4full weights profile
cometkim
committed
4b84dd5
feat(convert): nvfp4full encoder, calibration, converter, and verifier for qwen3.8-27b
cometkim
committed
546e2d0
docs(maintainer): define the qwen3.8-27b nvfp4full artifact contract
cometkim
committed
f78f31e
docs(maintainer): record nvfp4full gpqa-diamond 89.39% quality gate
cometkim
committed
f8ed361
docs(cards): add Qwen3.8-27B-nvfp4full model card
cometkim
committed
5eaa26d
docs(cards): point nvfp4full card at its published repository
cometkim
committed
ee45268
docs(cards): complete model-index source metadata
cometkim
committed
4b7cb23
docs(cards): point nvfp4full card at its published fork branch
cometkim
committed
7499426
docs(readme): list the published qwen3.8 nvfp4full artifact
cometkim
committed
f73419d
docs: compare gpqa against the official's republished 90.40% score
cometkim
committed
3a2a9e0

Commits on Aug 22, 2026

Commits on Aug 23, 2026

feat(frontend): add Jinja chat template override

Doelfke
authored and
Gevil
committed
41ef241
Merge pull request #1 from Doelfke/feat/jinja-chat-template

Doelfke
authored and
Gevil
committed
9991a30
fix(templates): honor reasoning effort in custom Jinja chat templates

Doelfke
authored and
Gevil
committed
a5ff456
docs: trim verbose custom-chat-template sections

Doelfke
authored and
Gevil
committed
5da8867
feat(serve): advertise max_model_len in the /v1/models contract

SolettaSolaris
authored and
Gevil
committed
e743098
fix(templates): distinguish break from continue in minja loop-control tokens

authored and
Gevil
committed
3d840f2
docs(serving): document max_model_len in models contract

SolettaSolaris
authored and
Gevil
committed
b40a7d2
fix(templates): compare minja JSON arrays/objects with native equality

authored and
Gevil
committed
7b40728
build: pin docker.io base images for podman registry resolution
Gevil
committed
ac9ee4d
Merge PR #43: Jinja chat template override (--chat-template-file)
Gevil
committed
b189de0
fix(frontend): derive starts_in_reasoning from the rendered prompt
Gevil
committed
fc4f6ca
Merge upstream/master (feaf4dd) into fork/nvfp4full-merged
Gevil
committed
0b9fc6e
Merge cometkim/ninfer feat/qwen3.8-nvfp4full (nvfp4full weights profile)
Gevil
committed
0e044c0
docs: personal-fork adoption record (ADOPTION.md) + README note
Gevil
committed
191c13d
merge upstream-PR#55 (local): feat(serve): report prefix-cache hits in the chat-completions usage
Gevil
committed
677b9e8
merge upstream-PR#67 (local): perf(ops): fold the moe router selection into the last D1 block
Gevil
committed
91b150b
merge upstream-PR#69 (local): perf(ops): prefetch shared expert down weights behind moe router
Gevil
committed
46ff4bd
merge upstream-PR#65 (local): fix(serve): type tool-call parameters by their declared schema
Gevil
committed
0e659f7
merge upstream-PR#57 (local): fix(serving): export parse_content_parts and unit-test the allowed_types allowlist
Gevil
committed
fdcf683
merge upstream-PR#61 (local): feat(serve): bound each image with a Vision-token budget
Gevil
committed
325986e
test(serve): cover usage.prompt_tokens_details.cached_tokens; fix three stale test bugs
Gevil
committed
ac76ad2
docs(adoption): Tier 1 done - verification summary + PR link
Gevil
committed
4eae6d1
serve: accept chat_template_kwargs.reasoning_effort (Sharp template steering)
Gevil
committed
0e498b5
merge pr-79 (local): fix(serve): decouple warmup from client-facing request deadline
Gevil
committed
f91b7c8
merge pr-54 (local): fix(serve): log the exception that terminates the process
Gevil
committed
9a9d096
docs(adoption): Tier 2 executed - NEW-A/NEW-B + #79/#54, #50 skip record, verification summary
Gevil
committed
ab63789
cherry-pick dylanbrodiefafard/experimental d85278b1 (local)
Gevil
committed
d48c9e6
cherry-pick dylanbrodiefafard/experimental 7374b27a (local)
Gevil
committed
b82f1ed
cherry-pick dylanbrodiefafard/experimental 0787d7cf (local)
Gevil
committed
03514cf
cherry-pick dylanbrodiefafard/experimental dd1bd651 (local)
Gevil
committed
ebe76d2
cherry-pick dylanbrodiefafard/experimental bc341b75 (local)
Gevil
committed
98bbc47
cherry-pick MichaelDementii/chunked-prefill 286627dd (local)
Gevil
committed
ceddaf2
cherry-pick MichaelDementii/decode-suite 10fc5bbd (local)
Gevil
committed
4002244
cherry-pick MichaelDementii/decode-suite f8ad432a (local)
Gevil
committed
32c0ded
cherry-pick dylanbrodiefafard/fix-tool-calls-checkpointing 673d695c (local)
Gevil
committed
281cf9e
fix(frontend): strip stray vision-pad markers instead of rejecting the request
Gevil
committed
22c46df

Commits on Aug 24, 2026