feat(ds4v): dflash vision serving runtime with bounded IQ85 conversion - #722

Merged

davide221 merged 131 commits into

Luce-Org:mainfrom

marcelormendes:ds4v/quant85

Sep 23, 2026

Commits on Sep 4, 2026

feat(ds4v): MIX converter for Vision-Exp safetensors plus codec test
marcelormendes
committed
af84f44
test(ds4v): reproduce collapsed fitter and repair clean-build test wiring
marcelormendes
committed
a53acc9
fix(ds4v): separate collapsed BF16 codebook levels and stamp repairs
marcelormendes
committed
9e403cd
test(ds4v): replay an individual source expert through calibration
marcelormendes
committed
07e3284
feat(ds4v): export lossless vision projector GGUF
marcelormendes
committed
fd5a994
chore(ds4v): trim exporter test commentary
marcelormendes
committed
ef64f62
feat(ds4v): implement isolated BF16 vision tower runtime and CPU probes
marcelormendes
committed
c563fbf
fix(ds4v): reserve measured scratch before graph allocation
marcelormendes
committed
dfa9719
test(ds4v): compare parent stages and reject malformed projectors
marcelormendes
committed
076de6e
fix(ds4v): pin diagnostic snapshots independently of tensor views
marcelormendes
committed
8758e09
test(ds4v): isolate block kernels with native inputs and BF16 ULP metrics
marcelormendes
committed
8bac576
test(ds4v): verify transactional reload and strict tensor table rejection
marcelormendes
committed
d26f1c9
test(ds4v): measure parent BF16 Math SDPA sensitivity
marcelormendes
committed
da3fa6e
test(ds4v): identify actual parent SDPA dispatch and parser rejection
marcelormendes
committed
1cf2b26
fix(ds4v): preserve parent Math SDPA operand scaling order
marcelormendes
committed
4573248
test(ds4v): cover multihead attention and BF16 halfway rounding
marcelormendes
committed
2bfba86
docs(ds4v): record CPU qualification and remaining numerical limits
marcelormendes
committed
a23d7cb
test(ds4v): isolate vendored GGUF reader import
marcelormendes
committed
8e28cfb
Revert "test(ds4v): isolate vendored GGUF reader import"
marcelormendes
committed
782bf15
test(ds4v): cover complete image token budget
marcelormendes
committed
5845ed5
fix(ds4v): bound the complete image block before encoding
marcelormendes
committed
5bf705e
test(ds4v): specify image routing and raw visibility policy
marcelormendes
committed
e30c1b0
feat(ds4v): add exact RGB vision preprocessing
marcelormendes
committed
a238e8e
feat(ds4v): add bounded image routing and raw visibility helpers
marcelormendes
committed
60ce979
test(ds4v): specify bounded image data URL transport
marcelormendes
committed
0306a3f
feat(ds4v): parse bounded image data URLs and preserve content order
marcelormendes
committed
e09f8da
docs(ds4v): record image policy contracts and CPU proof
marcelormendes
committed
5da9272
docs(ds4v): record image transport contract and proof
marcelormendes
committed
4acc20d
feat(ds4v): add bounded JPEG and PNG decode gate
marcelormendes
committed
1fa0a2e
build(ds4v): prepare explicit HIP tower qualification probe
marcelormendes
committed
127bb61
test(ds4v): bound image message nesting before JSON copy
marcelormendes
committed
2af9862
fix(ds4v): isolate libjpeg probe build
marcelormendes
committed
60af19f
fix(ds4v): validate message depth and placeholders before copying
marcelormendes
committed
25f6105
chore(ds4v): keep codec probe dependency focused
marcelormendes
committed
724e4ef
fix(ds4v): reuse server HIP compatibility definitions in probe
marcelormendes
committed
4bf7270
test(ds4v): cover RGB conversion and EXIF handling
marcelormendes
committed
6a2b381
test(ds4v): capture decoder review regressions
marcelormendes
committed
a28abfd
test(ds4v): specify transactional image prompt preparation
marcelormendes
committed
c24e823
fix(ds4v): bound PNG inflate and match source decoding
marcelormendes
committed
6d28845
feat(ds4v): prepare owned image prompts with checked expansion
marcelormendes
committed
a2489a4
test(ds4v): prove prompt ownership and document CPU qualification
marcelormendes
committed
859f2f7
docs(ds4v): correct LodePNG notice
marcelormendes
committed
edb3b0e
feat(ds4v): compose verified prompt and preprocessing units
marcelormendes
committed
52c6ba7
feat(ds4v): compose verified image transport
marcelormendes
committed
101f69c
feat(ds4v): compose verified image routing policy
marcelormendes
committed
20ae935
test(ds4v): specify CPU image preparation composition
marcelormendes
committed
1096fab
test(ds4v): link existing Jinja unicode helper in composition probe
marcelormendes
committed
a5aa8a3
feat(ds4v): compose accepted image preparation in CPU probe
marcelormendes
committed
15a10c5
test(ds4v): freeze bounded CPU preparation composition proof
marcelormendes
committed
0065158

Commits on Sep 5, 2026

Commits on Sep 6, 2026

Commits on Sep 10, 2026

Commits on Sep 21, 2026

Merge main (e0048e019) into ds4v/quant85

mrciffa
and
claude
committed
344713d
build(ds4v): make hipBLASLt optional for the HIP backend

mrciffa
and
claude
committed
94856ca
chore(ds4v): drop research logs and evidence receipts

mrciffa
and
claude
committed
6f784d4
fix(ds4v): keep the text path identical to main without --mmproj

mrciffa
and
claude
committed
7651512
chore(ds4v): drop the rejected IQ85 recipe

mrciffa
and
claude
committed
f248dfd
test(ds4v): run the image unit tests in the main build, drop standalone probes

mrciffa
and
claude
committed
212b3a6
fix(ggml): correct the RPC op-count guard and shorten the vendor note

mrciffa
and
claude
committed
c42b579
docs(ds4v): state the real verification status of image serving

mrciffa
and
claude
committed
fa7ff8f
fix(ds4v): scan for image markers only with a projector loaded; use model dimensions

mrciffa
and
claude
committed
633df96
fix(server): leave image parts alone on backends without image input

mrciffa
and
claude
committed
5c81edc
test(ds4): follow the generalised image-bias loader message

mrciffa
and
claude
committed
02b510f
fix(server): run image extraction and redaction only for an image-capable backend

mrciffa
and
claude
committed
89fd666
fix(ds4v): refuse --mmproj at startup when the build has no vision ops

mrciffa
and
claude
committed
f9cc9e3
fix(ds4v): ignore the MTP block's image bias in the loader check; drop the ratio-pattern test

mrciffa
and
claude
committed
37dfa4d
test(ds4): the loader no longer pins the DS4V layer count and vocabulary

mrciffa
and
claude
committed
7d4d7ac
fix(ds4v): check for the vision ops before loading weights; report the loader's reason

mrciffa
and
claude
committed
4196a70
build(ds4v): pass DOWNLOAD_EXTRACT_TIMESTAMP only on CMake >= 3.24

mrciffa
and
claude
committed
13a5dae
build(ds4v): build the MIX converter only when its source is in the tree

mrciffa
and
claude
committed
84591eb
refactor(vision): move the model-independent image pieces out of deepseek4

mrciffa
and
claude
committed
a6e4e72
feat(ds4v): load the image router bias from published GGUFs

mrciffa
and
claude
committed
57fe97d
feat(ds4): load llama.cpp-converted DeepSeek4 GGUFs without deepseek4.vocab_size

mrciffa
and
claude
committed
8544d76
fix(ds4v): count the GPU driver's page pool in the image memory admission

mrciffa
and
claude
committed
b91eea9
build(docker): build the DS4V vision ops in the ROCm image; keep the admission test machine independent

mrciffa
and
claude
committed
f023a8a
feat(ds4v): image input on one GPU holding the whole model

mrciffa
and
claude
committed
49cabab
fix(ds4v): review fixes - images outside the last prefill chunk, 16-bit PNGs, offline builds

mrciffa
and
claude
committed
b3409fc
Merge remote-tracking branch 'lucebox/main' into pr722-clean
mrciffa
committed
38e0cfd
build(vision): vendor lodepng instead of downloading it at configure time

mrciffa
and
claude
committed
12a0b29
refactor(vision): share the Pillow-exact resize, add a projector file reader

mrciffa
and
claude
committed
d1d5be7
feat(qwen35): image input for Qwen3.5 / Qwen3.8 through --mmproj

mrciffa
and
claude
committed
b52c981
Merge main: engine rename from dflash to luce

mrciffa
and
claude
committed
47e9259
fix: review findings in the page pool estimate and the MIX converter

mrciffa
and
claude
committed
cbf8691
fix(qwen35): write the request-visible vision config once, never snapshot an image prefill

mrciffa
and
claude
committed
f18750f
fix(vision): review findings in the Qwen image path, the projector reader and DS4V admission

mrciffa
and
claude
committed
8cca0e9
fix(vision): release the projector before its backend, smaller review items

mrciffa
and
claude
committed
27b08e7

Commits on Sep 22, 2026

perf(qwen35): fused attention in the vision tower

mrciffa
and
claude
committed
47e4703
docs(vision): Q8_0 projector measurement

mrciffa
and
claude
committed
901ef2d
perf(tools): quantize experts on every core in the MIX converter

mrciffa
and
claude
committed
c202867
fix(deepseek4): grouped output projection only for a quantized tensor

mrciffa
and
claude
committed
6e5a0ca
fix(tools): name the MIX file after its checkpoint

mrciffa
and
claude
committed
0a7ed37
docs(vision): DS4V on our own ROCMFP MIX conversion

mrciffa
and
claude
committed
61c07ff
feat(tools): write the MIX file with the shipped DeepSeek-V4-Flash recipe

mrciffa
and
claude
committed
6461136
feat(tools): per-expert importance and embedded codebooks in the MIX converter

mrciffa
and
claude
committed
0d349b8
docs(vision): DS4V MIX results with the shipped recipe

mrciffa
and
claude
committed
a098ecd
docs(vision): Qwen3.8 results with the Lucebox model file and a Q8_0 projector

mrciffa
and
claude
committed
25735a3
build(vision): fetch lodepng at a pinned commit instead of vendoring it

mrciffa
and
claude
committed
b35966e

Commits on Sep 23, 2026

docs(vision): quick start from the published Lucebox files, LUCE_MMPROJ for Docker

mrciffa
and
claude
committed
34d0b06