Tags

build: add local wheel-file and classic DLL build scripts

- build-wheel-file.bat mirrors build-wheel.bat (same kv-stream flags:
  GGML_CUDA + FA_ALL_QUANTS, sm_120a) but keeps the .whl in dist/
  instead of installing into the environment.
- build-kvstream.bat builds the classic shared-library layout
  (ggml/llama/llama-common/mtmd DLLs) for in-place llama_cpp/lib use.

vendor: pin kv-stream submodule to LLAI-lab mirror with graceful mmpr…

…oj OOM error

RaymondHuang210129's feature/adaptive-kv-stream branch was rebased, leaving
the release-pinned d873e5d lineage unreachable from any branch there. Host
that lineage (plus the clip.cpp graceful-OOM fix, 2f9dfd0) on the
LLAI-lab/llama.cpp fork so fresh clones can resolve the gitlink, and point
.gitmodules at it.

llama_cpp: support adaptive KV streaming (kv-stream fork)

Point vendor/llama.cpp at RaymondHuang210129/llama.cpp-adaptive-kv-streaming
(feature/adaptive-kv-stream, d873e5db9), which adds block-granular KV cache
streaming for CUDA so long contexts fit when weights leave too little VRAM.

Python-side adaptation:
- llama_context_params: add kv_stream_stage_mib; Llama() accepts the same
  kwarg and falls back to LLAMA_ARG_KV_STREAM_STAGE_MIB from the environment
- mtmd_context_params: drop the device field absent from the fork's mtmd.h
- mark upstream-only mtmd symbols (set_mergeable, get_placeholder,
  tokenize_from_parts, init_opt_default) as optional so import works
- guard lazy_mode for model params without that field

Bump version to 0.3.49

- 0.3.49 mainly expands the speculative decoding stack with DFlash2, DFlash, and DSpark support, while also improving MTMD video handling, cache recovery, runtime timing, and API compatibility.

Signed-off-by: JamePeng <jame_peng@sina.com>

Bump version to 0.3.49

- 0.3.49 mainly expands the speculative decoding stack with DFlash2, DFlash, and DSpark support, while also improving MTMD video handling, cache recovery, runtime timing, and API compatibility.

Signed-off-by: JamePeng <jame_peng@sina.com>

Bump version to 0.3.49

- 0.3.49 mainly expands the speculative decoding stack with DFlash2, DFlash, and DSpark support, while also improving MTMD video handling, cache recovery, runtime timing, and API compatibility.

Signed-off-by: JamePeng <jame_peng@sina.com>

Bump version to 0.3.49

- 0.3.49 mainly expands the speculative decoding stack with DFlash2, DFlash, and DSpark support, while also improving MTMD video handling, cache recovery, runtime timing, and API compatibility.

Signed-off-by: JamePeng <jame_peng@sina.com>

Bump version to 0.3.49

- 0.3.49 mainly expands the speculative decoding stack with DFlash2, DFlash, and DSpark support, while also improving MTMD video handling, cache recovery, runtime timing, and API compatibility.

Signed-off-by: JamePeng <jame_peng@sina.com>

Bump version to 0.3.49

- 0.3.49 mainly expands the speculative decoding stack with DFlash2, DFlash, and DSpark support, while also improving MTMD video handling, cache recovery, runtime timing, and API compatibility.

Signed-off-by: JamePeng <jame_peng@sina.com>

Bump version to 0.3.49

- 0.3.49 mainly expands the speculative decoding stack with DFlash2, DFlash, and DSpark support, while also improving MTMD video handling, cache recovery, runtime timing, and API compatibility.

Signed-off-by: JamePeng <jame_peng@sina.com>