v0.3.9-cu126-AVX2-win-20250702

Update Submodule vendor/llama.cpp bf2a99e..de56944

v0.3.9-cu126-AVX2-linux-20250701

Update Submodule vendor/llama.cpp bf2a99e..de56944

v0.3.9-cu124-AVX2-win-20250702

Update Submodule vendor/llama.cpp bf2a99e..de56944

v0.3.9-cu124-AVX2-linux-20250701

Update Submodule vendor/llama.cpp bf2a99e..de56944

v0.3.9-cu126-AVX2-win-20250525

Implement mtmd_cpp.py, base on tools/mtmd/mtmd.h #MTMD_API
Note: llava_cpp.py will be removed after llama_chat_format.py is adjusted.
It cannot connect to llava.dll (it is now mtmd.dll)

Sync kv-cache : add SWA support

Update llama.cpp API code 20250513

Sync context : remove logits_all flag and update API

Update LLAVA_API code in llava_cpp.py

Sync llava_cpp code: Update clip.h function API

Sync quantize: Handle user-defined quantization levels for additional tensors (#12511)

Sync llama : Support llama 4 text-only

Update llama : add option to override model tensor buffers

Sync llama-vocab : add SuperBPE pre-tokenizer

class LlamaSampler: append add_xtc(), add_top_n_sigma() and add_dry()

Update Submodule vendor/llama.cpp to f5cd27b

v0.3.9-cu126-AVX2-linux-20250525

Implement mtmd_cpp.py, base on tools/mtmd/mtmd.h #MTMD_API
Note: llava_cpp.py will be removed after llama_chat_format.py is adjusted.
It cannot connect to llava.dll (it is now mtmd.dll)

build-wheels-linux -DCMAKE_CUDA_ARCHITECTURES add 70-real;75-real

Sync kv-cache : add SWA support

Update llama.cpp API code 20250513

Sync context : remove logits_all flag and update API

Update LLAVA_API code in llava_cpp.py

Sync llava_cpp code: Update clip.h function API

Sync quantize: Handle user-defined quantization levels for additional tensors (#12511)

Sync llama : Support llama 4 text-only

Update llama : add option to override model tensor buffers

Sync llama-vocab : add SuperBPE pre-tokenizer

class LlamaSampler: append add_xtc(), add_top_n_sigma() and add_dry()

Update Submodule vendor/llama.cpp to f5cd27b

v0.3.9-cu124-AVX2-win-20250525

Implement mtmd_cpp.py, base on tools/mtmd/mtmd.h #MTMD_API
Note: llava_cpp.py will be removed after llama_chat_format.py is adjusted.
It cannot connect to llava.dll (it is now mtmd.dll)

Sync kv-cache : add SWA support

Update llama.cpp API code 20250513

Sync context : remove logits_all flag and update API

Update LLAVA_API code in llava_cpp.py

Sync llava_cpp code: Update clip.h function API

Sync quantize: Handle user-defined quantization levels for additional tensors (#12511)

Sync llama : Support llama 4 text-only

Update llama : add option to override model tensor buffers

Sync llama-vocab : add SuperBPE pre-tokenizer

class LlamaSampler: append add_xtc(), add_top_n_sigma() and add_dry()

Update Submodule vendor/llama.cpp to f5cd27b

v0.3.9-cu124-AVX2-linux-20250525

Implement mtmd_cpp.py, base on tools/mtmd/mtmd.h #MTMD_API
Note: llava_cpp.py will be removed after llama_chat_format.py is adjusted.
It cannot connect to llava.dll (it is now mtmd.dll)

build-wheels-linux -DCMAKE_CUDA_ARCHITECTURES add 70-real;75-real

Sync kv-cache : add SWA support

Update llama.cpp API code 20250513

Sync context : remove logits_all flag and update API

Update LLAVA_API code in llava_cpp.py

Sync llava_cpp code: Update clip.h function API

Sync quantize: Handle user-defined quantization levels for additional tensors (#12511)

Sync llama : Support llama 4 text-only

Update llama : add option to override model tensor buffers

Sync llama-vocab : add SuperBPE pre-tokenizer

class LlamaSampler: append add_xtc(), add_top_n_sigma() and add_dry()

Update Submodule vendor/llama.cpp to f5cd27b

v0.3.8-cu126-AVX2-win-20250506

0.3.8_Final_version

v0.3.8-cu126-AVX2-linux-20250506

0.3.8_Final_version