Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

mtmd: support DeepSeek-V4-Flash-Vision-Exp (#28133)

Browse Source 
* mtmd: support DeepSeek-V4-Flash-Vision-Exp

* handle min/max token counts from CLI

* rm debugging

* use GGML_ROPE_TYPE_VISION

* nits

* apply review comments

* correct token count
This commit is contained in:
Xuan-Son Nguyen
2026-09-02 16:43:43 +02:00
committed by GitHub
parent 8e93a9773b
commit 7798007a29
13 changed files with 504 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gguf-py/gguf/constants.py
+10
View File
@@ -950,6 +950,9 @@ class MODEL_TENSOR(IntEnum):
V_RESMPL_PROJ = auto() # minicpmv
V_RESMPL_QUERY = auto() # minicpmv
V_TOK_EMBD_IMG_BREAK = auto() # pixtral
V_TOK_EMBD_IMG_START = auto() # deepseek4v
V_TOK_EMBD_IMG_END = auto() # deepseek4v
V_TOK_EMBD_IMG_PAD = auto() # deepseek4v
V_MM_PATCH_MERGER = auto() # mistral small 3.1
V_DS_NORM = auto() # qwen3vl
V_DS_FC1 = auto() # qwen3vl
@@ -1696,6 +1699,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.V_RESMPL_PROJ: "resampler.proj",
MODEL_TENSOR.V_RESMPL_QUERY: "resampler.query",
MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: "v.token_embd.img_break", # pixtral
MODEL_TENSOR.V_TOK_EMBD_IMG_START: "v.token_embd.img_start", # deepseek4v
MODEL_TENSOR.V_TOK_EMBD_IMG_END: "v.token_embd.img_end", # deepseek4v
MODEL_TENSOR.V_TOK_EMBD_IMG_PAD: "v.token_embd.img_pad", # deepseek4v
MODEL_TENSOR.V_MM_PATCH_MERGER: "mm.patch_merger", # mistral small 3.1
MODEL_TENSOR.V_DS_NORM: "v.deepstack.{bid}.norm",
MODEL_TENSOR.V_DS_FC1: "v.deepstack.{bid}.fc1",
@@ -2030,6 +2036,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.V_RESMPL_PROJ,
MODEL_TENSOR.V_RESMPL_QUERY,
MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
MODEL_TENSOR.V_TOK_EMBD_IMG_START,
MODEL_TENSOR.V_TOK_EMBD_IMG_END,
MODEL_TENSOR.V_TOK_EMBD_IMG_PAD,
MODEL_TENSOR.V_MM_PATCH_MERGER,
MODEL_TENSOR.V_MM_MERGER_FC1,
MODEL_TENSOR.V_MM_MERGER_FC2,
@@ -5645,6 +5654,7 @@ class VisionProjectorType:
DOTS3NOTE_A = "dots3note_a" # audio
DEEPSEEKOCR = "deepseekocr"
DEEPSEEKOCR2 = "deepseekocr2"
DEEPSEEK4V = "deepseek4v"
LFM2A = "lfm2a" # audio
MUSIC_FLAMINGO = "musicflamingo" # audio
GLM4V = "glm4v"