feat(MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm): run the DeepSeek-V4-Flash-Vision tower on CUDA against llama.cpp b10766

.agents/claims/CLAIM-DEEPSEEK-V4-VISION-W1.md

@@ -0,0 +1,5 @@
+# CLAIM-DEEPSEEK-V4-VISION-W1
+
+| Claim | Row IDs | Agent | Worktree / remote dir | Branch | Owned scope | State | Last update |
+|---|---|---|---|---|---|---|---|
+| `CLAIM-DEEPSEEK-V4-VISION-W1` | `MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm` (`ACTIVE`) | OpenAI Codex (gpt-5.6-sol), helper role, fresh retry implementer for issue [#2411](https://github.com/mudler/vllm.cpp/issues/2411) | `/home/mudler/_git/vllm.cpp/.claude/worktrees/deepseek-v4-vision-w1` | `row/MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm-W1` | Owns only W1 and these files: `include/vllm/multimodal/deepseek_v4_processor.h`; `src/vllm/multimodal/deepseek_v4_processor.cpp`; `tests/vllm/multimodal/test_deepseek_v4_encoding.cpp`; `tests/vllm/multimodal/test_deepseek_v4_image_processor.cpp`; `CMakeLists.txt`; `tests/CMakeLists.txt`; `.agents/model-matrix.md`; `.agents/specs/deepseek-v4-flash-vision.md`; and this claim. Excludes ViT, aligner, weights and loader, model forward, image-span visibility, public ABI, server, GPU work, large downloads, push, and merge. | `ACTIVE` | 2026-08-31 — W1 prompt encoding and RGB image preprocessing passed retry verification. Next: fresh scoped review, then the operator reruns the W1 gate before integration; W2 owns the ViT and aligner. |

.agents/claims/CLAIM-DEEPSEEK-V4-VISION-W2.md

@@ -0,0 +1,5 @@
+# CLAIM-DEEPSEEK-V4-VISION-W2
+
+| Claim | Row IDs | Agent | Worktree / remote dir | Branch | Owned scope | State | Last update |
+|---|---|---|---|---|---|---|---|
+| `CLAIM-DEEPSEEK-V4-VISION-W2` | `MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm` (`ACTIVE`) | OpenAI Codex (gpt-5.6-sol), helper role, fresh implementer for issue [#2411](https://github.com/mudler/vllm.cpp/issues/2411) | `/home/mudler/_git/vllm.cpp/.claude/worktrees/deepseek-v4-vision-w2` | `row/MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm-W2` | Owns only W2 and these files: `include/vllm/model_executor/models/deepseek_v4_vision.h`; `src/vllm/model_executor/models/deepseek_v4_vision.cpp`; `tests/vllm/models/test_deepseek_v4_vision.cpp`; `scripts/gen-deepseek-v4-vision-goldens.py`; `tests/parity/goldens/deepseek_v4_vision/`; `CMakeLists.txt`; `tests/CMakeLists.txt`; `.agents/specs/deepseek-v4-flash-vision.md` W2 evidence, Owed and Now; and this claim. Excludes DeepSeek-V4 text config, weights and loader, registered forward, image-span visibility, public ABI, server, device-specific kernels, GPU work, large downloads, push and merge. | `ACTIVE` | 2026-09-01 — W2 starts RED-first from the committed specification. The standalone ViT and aligner remain unreachable from production; W4 owns registered-forward reachability. |

.agents/claims/CLAIM-DEEPSEEK-V4-VISION-W3A.md

@@ -0,0 +1,5 @@
+# CLAIM-DEEPSEEK-V4-VISION-W3A
+
+| Claim | Row IDs | Agent | Worktree / remote dir | Branch | Owned scope | State | Last update |
+|---|---|---|---|---|---|---|---|
+| `CLAIM-DEEPSEEK-V4-VISION-W3A` | `MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm` (`ACTIVE`) | Anthropic Claude (claude-opus-5-1m), helper role, fresh implementer for issue [#2411](https://github.com/mudler/vllm.cpp/issues/2411) | `/home/mudler/_git/vllm.cpp/.wt/dsv4v-w3a` | `row/MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm-W3A` | Owns only W3A and these files: `include/vllm/model_executor/models/clip_mmproj_gguf.h`; `src/vllm/model_executor/models/clip_mmproj_gguf.cpp`; `tests/vllm/models/test_deepseek_v4_mmproj.cpp`; `tests/CMakeLists.txt`; `.agents/specs/deepseek-v4-flash-vision.md` W3A evidence and Owed; and this claim. Excludes the language-model GGUF arm and `deepseek_v4_weights.cpp` (W3B owns them), the Qwen3-VL projector path's behaviour, registered forward, image-span visibility, public ABI, server, device-specific kernels, GPU work, large downloads, push and merge. | `ACTIVE` | 2026-09-05 — W3A starts RED-first from the committed specification. The `deepseek4v` mmproj reader is unreachable from production; W4 owns registered-forward reachability and the spec lists the gap under `## Owed`. |

.agents/claims/CLAIM-DEEPSEEK-V4-VISION-W3B.md

@@ -0,0 +1,5 @@
+# CLAIM-DEEPSEEK-V4-VISION-W3B
+
+| Claim | Row IDs | Agent | Worktree / remote dir | Branch | Owned scope | State | Last update |
+|---|---|---|---|---|---|---|---|
+| `CLAIM-DEEPSEEK-V4-VISION-W3B` | `MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm` (`ACTIVE`) | Anthropic Claude (claude-opus-5-1m), helper role, fresh implementer for issue [#2411](https://github.com/mudler/vllm.cpp/issues/2411) | `/home/mudler/_git/vllm.cpp/.wt/dsv4v-w3b` | `row/MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm-W3B` | Owns only W3B, the `exp_probs_b_vl` routing bias's ACCOUNTING and LOADING, and these files: `include/vllm/model_executor/models/deepseek_v4.h` (the two layer-weight fields); `src/vllm/model_executor/models/deepseek_v4_weights.cpp`; `tests/vllm/models/test_deepseek_v4_mm_loader.cpp`; `tests/CMakeLists.txt`; `.agents/specs/deepseek-v4-flash-vision.md` W3B evidence and Owed; and this claim. Excludes per-token bias selection and the hash-layer replacement in `deepseek_v4_moe.cpp`, the non-causal image-span window, `clip_mmproj_gguf.{h,cpp}` and `deepseek_v4_vision.{h,cpp}` (W3A owns those), the registered forward, the public ABI, the server, device kernels, GPU work, large downloads, push and merge. | `ACTIVE` | 2026-09-05 — W3B starts RED-first from the committed specification. The loaded vision bias is a staged slice: nothing selects it, W4 owns the wiring, and the spec's `## Owed` lists it. |