mtmd : merge llava, gemma3 and minicpmv CLI into single llama-mtmd-cli
#13012
Conversation
ngxson
changed the title
llava-cli and gemma3-cli into single mtmd-clillama-mtmd-cli
| } | ||
|
|
||
| if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) { | ||
| // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode() |
There was a problem hiding this comment.
On second thought, it may not be a good idea to support real batching in clip_image_encode because the memory usage can blow up very quickly (some models use ~4k patches per image)
What we can do in the short term is to allow clip_image_encode to run the decode multiple times, i.e. simply copy this loop into clip_image_encode - this can be done in another PR
There was a problem hiding this comment.
I expect that in practice batching image encodes will have little to no utility since a single image is enough to typically saturate all the compute that is available. And as you mentioned, the memory for the compute buffers will grow very fast.
What we can do in the short term is to allow clip_image_encode to run the decode multiple times, i.e. simply copy this loop into clip_image_encode - this can be done in another PR
Sounds good.
| image_tokens->nx = clip_n_patches(ctx->ctx_clip); | ||
| image_tokens->ny = 1; |
There was a problem hiding this comment.
the ny - number of tokens in y direction is only used by qwen2vl for now (which we don't support yet in mtmd)
I should refactor mtmd_image_tokens very soon (in another PR). Otherwise, having both nx, ny for models not using M-RoPE looks very weird
There was a problem hiding this comment.
@ggerganov @slaren Tagging you here for visibility. Regarding to the discussion of tracking n_past internally by libllama. For qwen models, the first dimension of the 4D pos is the traditional n_past, but the rest are relative position (in the case of image, they are 2D X/Y coordinates)
So I think when refactoring llama_batch_ext, we should find a way to allow user to specify this "additional" dimensions:
(Note: for qwen, dimensions of each token in llama.cpp is currently [n_past, x, y, unused] - the unused dim is reserved for future usage, for ex. 3D spatial understanding)
ngxson
changed the title
llama-mtmd-clillama-mtmd-cli
| } | ||
|
|
||
| if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) { | ||
| // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode() |
There was a problem hiding this comment.
I expect that in practice batching image encodes will have little to no utility since a single image is enough to typically saturate all the compute that is available. And as you mentioned, the memory for the compute buffers will grow very fast.
What we can do in the short term is to allow clip_image_encode to run the decode multiple times, i.e. simply copy this loop into clip_image_encode - this can be done in another PR
Sounds good.
Co-authored-by: Georgi Gerganov <[email protected]>
…li` (ggml-org#13012) * mtmd : merge `llava-cli` and `gemma3-cli` into single `mtmd-cli` * support for minicpmv * remove cpp files of llava and minicpmv * update hot topics * mtmd : add not supported msg for qwen2vl * Update examples/llava/mtmd.cpp Co-authored-by: Georgi Gerganov <[email protected]> --------- Co-authored-by: Georgi Gerganov <[email protected]>
…li` (ggml-org#13012) * mtmd : merge `llava-cli` and `gemma3-cli` into single `mtmd-cli` * support for minicpmv * remove cpp files of llava and minicpmv * update hot topics * mtmd : add not supported msg for qwen2vl * Update examples/llava/mtmd.cpp Co-authored-by: Georgi Gerganov <[email protected]> --------- Co-authored-by: Georgi Gerganov <[email protected]>
This PR unifies all vision models supported by llama.cpp into
libmtmdQwen2VL is not yet merged for now, due to some complications in M-RoPE. This can be resolved after the
llama_batch_extrefactoring: #11875These models are supported:
llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF # and other sizes: 12b, 27b (+ QAT version) llama-mtmd-cli -hf guinmoon/MobileVLM-3B-GGUF --chat-template deepseek llama-mtmd-cli -hf THUDM/glm-edge-v-5b-gguf llama-mtmd-cli -hf second-state/Llava-v1.5-7B-GGUF --chat-template vicuna llama-mtmd-cli -hf cjpais/llava-1.6-mistral-7b-gguf --chat-template vicuna llama-mtmd-cli -hf ibm-research/granite-vision-3.2-2b-GGUF llama-mtmd-cli -hf second-state/MiniCPM-Llama3-V-2_5-GGUF llama-mtmd-cli -hf openbmb/MiniCPM-V-2_6-gguf llama-mtmd-cli -hf openbmb/MiniCPM-o-2_6-ggufNOTE:
Yi-VL-6Bis removed from the test because:Follow-up PRs:
clip.cpp(used by glm-edge), they should be processed as text tokens ==> will be a breaking changeREADMEfiles)