Skip to content

Help wanted: test DeepSeek V4 Flash Vision-Exp AProjQ4 on Metal, CUDA, and ROCm #929

Description

@GiorgioOppo

Summary

We are looking for volunteers to test the new DeepSeek V4 Flash Vision-Exp AProjQ4 GGUF on:

  • Apple Silicon / Metal
  • NVIDIA / CUDA
  • AMD Strix Halo / ROCm
  • Fully resident and SSD-streaming configurations

The runtime implementation is tracked in PR #621.

This request complements #838, which covers the non-Vision AProjQ4 model, and follows the original Vision support request in #851.

You only need to test one hardware/backend combination. Reports of crashes, incorrect image understanding, quality drift, unexpectedly low performance, or successful runs are all useful.

Model artifact

The experimental model is available from Hugging Face PR #22:

DeepSeek-V4-Flash-Vision-Exp-IQ2XXS-w2Q2K-AProjQ4-SExpQ8-OutQ8-imatrix.gguf

Size:    84,420,584,608 bytes (78.62 GiB)
SHA-256: a1d0fde967c910e37e4bea00c4db1b2e8a88978770828a95772bc66837386cb8

The model was produced by requantizing 215 dense-attention projections—five projections across 43 layers—from Q8_0 to Q4_K, using the 220k routed-and-dense DS4 imatrix.

All tensors outside the selected attention projections retain the source model’s quantization layout.

The matching Vision-Exp encoder sidecar is also required:

DeepSeek-V4-Flash-Vision-Encoder.gguf

Download

Install the current Hugging Face CLI if needed, then run:

mkdir -p gguf

hf download antirez/deepseek-v4-gguf \
  DeepSeek-V4-Flash-Vision-Exp-IQ2XXS-w2Q2K-AProjQ4-SExpQ8-OutQ8-imatrix.gguf \
  --revision refs/pr/22 \
  --local-dir gguf

hf download antirez/deepseek-v4-gguf \
  DeepSeek-V4-Flash-Vision-Encoder.gguf \
  --revision main \
  --local-dir gguf

Verify the language-model GGUF:

shasum -a 256 \
  gguf/DeepSeek-V4-Flash-Vision-Exp-IQ2XXS-w2Q2K-AProjQ4-SExpQ8-OutQ8-imatrix.gguf

On Linux, use sha256sum instead.

Build the current PR head

Please fetch the latest PR head immediately before testing and report its exact commit:

git clone https://github.com/antirez/ds4.git ds4-vision-q4
cd ds4-vision-q4

git fetch origin pull/621/head:pr-621
git checkout pr-621
git rev-parse HEAD

Build for your backend:

# Apple Silicon / Metal
make clean
make

# DGX Spark / GB10
make clean
make cuda-spark

# Other NVIDIA CUDA systems
make clean
make cuda-generic

# AMD Strix Halo / ROCm
make clean
make strix-halo -j"$(nproc)"

Inspect the model first

MODEL=gguf/DeepSeek-V4-Flash-Vision-Exp-IQ2XXS-w2Q2K-AProjQ4-SExpQ8-OutQ8-imatrix.gguf
VISION=gguf/DeepSeek-V4-Flash-Vision-Encoder.gguf

./ds4 --cpu --inspect -m "$MODEL"

Expected high-level metadata:

GGUF v3
1328 tensors
43 layers
q4_k: 215 tensors
q8_0: 130 tensors
file size: 78.62 GiB

Run the model

Metal with SSD streaming

./ds4 --metal --ssd-streaming \
  --ctx 4096 --temp 0 \
  -m "$MODEL" \
  --vision "$VISION"

CUDA with SSD streaming

./ds4 --cuda --ssd-streaming \
  --ctx 4096 --temp 0 \
  -m "$MODEL" \
  --vision "$VISION"

ROCm with SSD streaming

./ds4 --rocm --ssd-streaming \
  --ctx 4096 --temp 0 \
  -m "$MODEL" \
  --vision "$VISION"

For fully resident testing, remove --ssd-streaming. The model, vision encoder, KV cache, and runtime buffers must all fit in available memory.

Inside the interactive CLI, submit a PNG or JPEG with:

/read /path/to/test-image.png

Please use public or non-sensitive images.

Minimum useful test

Please try at least:

  1. One ordinary photograph with several recognizable objects.
  2. One screenshot or image containing readable text.
  3. One diagram, chart, or spatial-reasoning image.
  4. A follow-up question referring to details from the same image.
  5. A text-only prompt in the same session, to detect general language-quality regressions.

Check for:

  • crashes or backend errors;
  • NaNs or non-finite logits;
  • repeated BOS tokens or looping output;
  • incorrect image-token handling;
  • severe hallucinations or loss of image grounding;
  • unexpected quality differences from the AProjQ8 model;
  • unusually slow vision encoding, prefill, or generation;
  • excessive memory usage or SSD-cache churn.

Optional AProjQ8 comparison

If storage and memory permit, compare against the matching Vision-Exp AProjQ8 language model:

DeepSeek-V4-Flash-Vision-Exp-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8.gguf

Please use the same:

  • DS4 commit;
  • backend and execution mode;
  • context size;
  • vision encoder;
  • image;
  • prompt;
  • temperature.

A matched AProjQ8/AProjQ4 comparison is especially valuable for detecting quantization-related quality drift.

Reporting template

Please post results using this template:

Hardware:
RAM / VRAM:
OS:
Backend:
Build command:
DS4 commit:
Resident or SSD streaming:
Context size:

Q4 SHA-256:
Vision encoder filename:

Exact command:
Image format and dimensions:
Image or public source link:

Result:
Response excerpt:
Vision grounding issues:
Crash/error output:

Vision/prefill timing:
Generation speed:
Peak or reported memory:

AProjQ8 comparison, if available:
Additional environment variables:
Other notes:

Please include complete error messages where possible, but redact private paths, tokens, credentials, and sensitive image content.

Current validation status

The GGUF has already passed:

  • structural GGUF inspection;
  • tensor-count and tensor-type validation;
  • strict imatrix compatibility checks;
  • a text-only Metal SSD-streaming smoke test covering all 43 layers.

Full multimodal validation across different GPUs and operating systems is still needed. No broad Vision quality-equivalence claim is being made yet.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions