Multimodal embedding (tested on Qwen-VL-Embedding)

roj234 · roj234 · commit 07a71ae2bd25 · 2026-02-21T22:19:50.000+08:00
diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py
@@ -672,7 +672,12 @@ def reset(self):
         if self.batch is not None:
             self.batch.n_tokens = 0
 
-    def set_batch(self, batch: Sequence[int], n_past: llama_cpp.llama_pos, logits_all: bool):
+    def set_batch(self,
+        batch: Sequence[int],
+        n_past: llama_cpp.llama_pos,
+        logits_all: bool,
+        logits_last: bool = True
+    ):
         if len(batch) > self.n_tokens_capacity:
              raise IndexError(f"Input batch size {len(batch)} exceeds capacity {self.n_tokens_capacity}")
 
@@ -684,7 +689,7 @@ def set_batch(self, batch: Sequence[int], n_past: llama_cpp.llama_pos, logits_al
             self.batch.seq_id[i][0] = 0
             self.batch.n_seq_id[i] = 1
             self.batch.logits[i] = logits_all
-        self.batch.logits[n_tokens - 1] = True
+        self.batch.logits[n_tokens - 1] = logits_last
 
     def add_sequence(self, batch: Sequence[int], seq_id: int, logits_all: bool):
         n_tokens = len(batch)
diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py
@@ -58,6 +58,9 @@
 from ._logger import set_verbose
 from ._utils import suppress_stdout_stderr
 
+from .mtmd_cpp import mtmd_context_params_default, mtmd_init_from_file
+from .mtmd import MultiModalContext
+
 
 class Llama:
     """High-level Python wrapper for a llama.cpp model."""
@@ -130,6 +133,10 @@ def __init__(
         # Misc
         spm_infill: bool = False,
         verbose: bool = True,
+        mmproj_path: str = None,
+        mmproj_use_gpu: Optional[bool] = None,
+        image_min_tokens: int = -1,
+        image_max_tokens: int = -1,
         # Extra Params
         **kwargs,  # type: ignore
     ):
@@ -426,6 +433,29 @@ def __init__(
             )
         )
 
+        if mmproj_path != None:
+            mparams = mtmd_context_params_default();
+            mparams.use_gpu          = mmproj_use_gpu if mmproj_use_gpu != None else n_gpu_layers == -1
+            mparams.print_timings    = verbose
+            mparams.n_threads        = self.n_threads
+            mparams.flash_attn_type  = self.context_params.flash_attn_type
+            mparams.warmup           = True
+            if image_min_tokens > 0:
+                mparams.image_min_tokens = image_min_tokens
+            if image_max_tokens > 0:
+                mparams.image_max_tokens = image_max_tokens
+
+            with suppress_stdout_stderr(disable=verbose):
+                mctx = mtmd_init_from_file(mmproj_path.encode("utf-8"), self._model.model, mparams)
+            if mctx is None:
+                raise RuntimeError(f"failed to load multimodal projection '{mmproj_path}'")
+
+            self.mtmd_context = self._stack.enter_context(
+                contextlib.closing(
+                    MultiModalContext(mctx)
+                )
+            )
+
         # Check for Encoder-Decoder architecture
         self._has_encoder = self._model.has_encoder()
         self._has_decoder = self._model.has_decoder()
diff --git a/llama_cpp/llama_embedding.py b/llama_cpp/llama_embedding.py
@@ -12,6 +12,8 @@
     LLAMA_POOLING_TYPE_LAST,
     LLAMA_POOLING_TYPE_RANK, # Specifically for Reranking models
 )
+from .mtmd import MediaChunk, mtmd_tokenize, mtmd_prefill
+from ._utils import suppress_stdout_stderr
 
 # Normalization modes for embedding vectors
 # See: https://github.com/ggml-org/llama.cpp/tree/master/examples/embedding#--embd-normalize-integer
@@ -168,7 +170,7 @@ def embed(
         if self.verbose:
             llama_cpp.llama_perf_context_reset(ctx)
         self._batch.reset()
-        llama_cpp.llama_memory_clear(llama_cpp.llama_get_memory(ctx), True)
+        llama_cpp.llama_memory_clear(llama_cpp.llama_get_memory(ctx), False)
 
         # Initialize State Variables
         results: List[Any] = []
@@ -219,7 +221,7 @@ def _decode_batch():
                         results.append(data)
 
             self._batch.reset()
-            llama_cpp.llama_memory_clear(llama_cpp.llama_get_memory(ctx), True)
+            llama_cpp.llama_memory_clear(llama_cpp.llama_get_memory(ctx), False)
             batch_seq_lens = []
 
         # Main Streaming Loop
@@ -427,3 +429,68 @@ def create_embedding(
                     print(f"Warning: Failed to calculate similarity matrix: {e}")
 
         return response
+
+
+    def embed_multimodal(
+        self,
+        prompt: str,
+        files: List[bytes | str] = [],
+
+        normalize: int = NORM_MODE_EUCLIDEAN,
+        return_count: bool = False,
+    ) -> Union[List[float], List[List[float]], Tuple[Any, int]]:
+
+        ctx = self._ctx.ctx
+        mctx = self.mtmd_context.ctx
+
+        # Determine if it is in Rerank mode
+        try:
+            pooling_type = self.pooling_type()
+        except AttributeError:
+            pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED
+        is_rank = (pooling_type == LLAMA_POOLING_TYPE_RANK)
+        is_none = (pooling_type == LLAMA_POOLING_TYPE_NONE) # Token-level embedding
+
+        out_dim = self.n_embd()
+
+        if self.verbose:
+            type_str = "TOKEN (None)" if is_none else ("RANK (Score)" if is_rank else "SEQ (Vector)")
+            print(f"LlamaEmbedding Debug: Mode={type_str} | Pooling={pooling_type} | Dim={out_dim}")
+
+        # Reset Context and Batch
+        if self.verbose:
+            llama_cpp.llama_perf_context_reset(ctx)
+        self._batch.reset()
+        llama_cpp.llama_memory_clear(llama_cpp.llama_get_memory(ctx), False)
+
+        # Initialize State Variables
+        result: Any = None
+
+
+        with suppress_stdout_stderr(disable=self.verbose):
+            tokens: MultimodalTokenList = mtmd_tokenize(mctx, prompt, files)
+
+        n_tokens = len(tokens)
+
+        if n_tokens == 0:
+            result = []
+        else:
+            n_past = mtmd_prefill(self._ctx, mctx, self._batch, tokens)
+
+            # Extract Embeddings
+            ptr = llama_cpp.llama_get_embeddings_ith(ctx, self._batch.n_tokens() - 1)
+            data = ptr[:out_dim]
+            data = self._normalize_vector(data, normalize)
+
+            result = data
+
+            self._batch.reset()
+            llama_cpp.llama_memory_clear(llama_cpp.llama_get_memory(ctx), False)
+
+        if self.verbose:
+            llama_cpp.llama_perf_context_print(ctx)
+
+        if return_count:
+            return result, n_tokens
+
+        return result
diff --git a/llama_cpp/mtmd.py b/llama_cpp/mtmd.py
@@ -0,0 +1,202 @@
+import llama_cpp
+from llama_cpp import LLAMA_TOKEN_NULL
+
+import llama_cpp.mtmd_cpp as mtmd
+from .mtmd_cpp import mtmd_input_chunk_type, mtmd_free
+from ._internals import LlamaContext, LlamaBatch
+
+import ctypes
+from typing import Union, List
+
+class TextChunk:
+    def __init__(self, tokens: List[int]):
+        self.tokens = tokens
+        self.n_tokens = len(tokens)
+
+class MediaChunk:
+    def __init__(self, chunk_ptr: ctypes.c_void_p):
+        self.chunk_ptr = mtmd.mtmd_input_chunk_copy(chunk_ptr)
+        self.n_tokens = mtmd.mtmd_input_chunk_get_n_tokens(self.chunk_ptr)
+
+    def __del__(self):
+        if self.chunk_ptr:
+            mtmd.mtmd_input_chunk_free(self.chunk_ptr)
+
+class MultimodalTokenList:
+    def __init__(self):
+        self.chunks: List[Union[TextChunk, MediaChunk]] = []
+        self.total_tokens = 0
+
+    def add(self, chunk_ptr: mtmd.mtmd_input_chunk_p_ctypes):
+        chunk_type = mtmd.mtmd_input_chunk_get_type(chunk_ptr)
+
+        if chunk_type in [
+            mtmd_input_chunk_type.MTMD_INPUT_CHUNK_TYPE_IMAGE,
+            mtmd_input_chunk_type.MTMD_INPUT_CHUNK_TYPE_AUDIO
+        ]:
+            m_chunk = MediaChunk(chunk_ptr)
+            self.chunks.append(m_chunk)
+            self.total_tokens += m_chunk.n_tokens
+
+        elif chunk_type == mtmd_input_chunk_type.MTMD_INPUT_CHUNK_TYPE_TEXT:
+            n_tokens_ref = ctypes.c_size_t()
+            text_tokens_ptr = mtmd.mtmd_input_chunk_get_tokens_text(chunk_ptr, ctypes.byref(n_tokens_ref))
+            tokens = [text_tokens_ptr[j] for j in range(n_tokens_ref.value)]
+            self.add_text(tokens)
+
+        else:
+            raise ValueError(f"Invalid chunk type {chunk_type}")
+
+    def add_text(self, tokens: List[int]):
+        if not tokens: return
+        # combine text nodes
+        if self.chunks and isinstance(self.chunks[-1], TextChunk):
+            self.chunks[-1].tokens.extend(tokens)
+            self.chunks[-1].n_tokens += len(tokens)
+        else:
+            self.chunks.append(TextChunk(tokens))
+        self.total_tokens += len(tokens)
+
+    def __len__(self):
+        return self.total_tokens
+
+
+class MultiModalContext:
+    def __init__(
+            self,
+            ctx
+    ):
+        self.ctx = ctx
+
+    def close(self):
+        if self.ctx is None:
+            return
+        mtmd_free(self.ctx)
+        self.ctx = None
+
+    def __del__(self):
+        self.close()
+
+
+# Simple FNV-1a hash implementation to match fnv_hash in C++
+def fnv_hash(data: bytes) -> str:
+    h = 0x811c9dc5
+    for b in data:
+        h = (h ^ b) * 0x01000193
+        h &= 0xffffffff
+    return f"{h:08x}"
+
+def mtmd_tokenize(
+    mctx: mtmd.mtmd_context_p,
+    prompt: str,
+    files_data: list[bytes | str]) -> MultimodalTokenList:
+
+    bitmaps = []
+    do_hash = False
+
+    for data in files_data:
+
+        bmp = None
+        if isinstance(data, str):
+            bmp = mtmd.mtmd_helper_bitmap_init_from_file(mctx, data.encode("utf-8"))
+        elif isinstance(data, bytes):
+            buf = (ctypes.c_ubyte * len(data)).from_buffer_copy(data)
+            bmp = mtmd.mtmd_helper_bitmap_init_from_buf(mctx, buf, len(buf))
+        elif isinstance(data, bytearray):
+            buf = (ctypes.c_ubyte * len(data)).from_buffer(data)
+            bmp = mtmd.mtmd_helper_bitmap_init_from_buf(mctx, buf, len(buf))
+
+        if bmp is None:
+            raise RuntimeError("Failed to load image or audio file")
+
+        if do_hash:
+            data_ptr = mtmd.mtmd_bitmap_get_data(bmp)
+            data_size = mtmd.mtmd_bitmap_get_n_bytes(bmp)
+
+            raw_node_data = ctypes.string_at(data_ptr, data_size)
+            h = fnv_hash(raw_node_data)
+            mtmd.mtmd_bitmap_set_id(bmp, h.encode('utf-8'))
+
+        bitmaps.append(bmp)
+
+    inp_txt = mtmd.mtmd_input_text(
+        text=prompt.encode('utf-8'),
+        add_special=True,
+        parse_special=True
+    )
+
+    chunks_ptr = mtmd.mtmd_input_chunks_init()
+
+    n_bitmaps = len(bitmaps)
+    if n_bitmaps > 0:
+        BitmapPtr = mtmd.mtmd_bitmap_p_ctypes * n_bitmaps
+        bitmaps_ptr = BitmapPtr(*bitmaps)
+    else:
+        bitmaps_ptr = None
+
+    res = mtmd.mtmd_tokenize(
+        mctx,
+        chunks_ptr,
+        ctypes.pointer(inp_txt),
+        bitmaps_ptr,
+        n_bitmaps
+    )
+
+    # TODO Hash based cache
+    for data in bitmaps:
+        mtmd.mtmd_bitmap_free(bmp)
+
+    if res != 0:
+        mtmd.mtmd_input_chunks_free(chunks_ptr)
+        raise RuntimeError(f"Tokenization failed with code {res}")
+
+    st = MultimodalTokenList()
+
+    n_chunks = mtmd.mtmd_input_chunks_size(chunks_ptr)
+    for i in range(n_chunks):
+        chunk_ptr = mtmd.mtmd_input_chunks_get(chunks_ptr, i)
+        st.add(chunk_ptr)
+
+    mtmd.mtmd_input_chunks_free(chunks_ptr)
+
+    return st
+
+def mtmd_prefill(
+    ctx: LlamaContext,
+    mctx: mtmd.mtmd_context_p,
+    batch: LlamaBatch,
+    mtmd_tokens: MultimodalTokenList
+) -> int:
+    n_past = 0
+    n_batch = ctx.n_batch()
+    total_chunks = len(mtmd_tokens.chunks)
+
+    for i, chunk in enumerate(mtmd_tokens.chunks):
+        is_last_chunk = (i == total_chunks - 1)
+
+        if isinstance(chunk, TextChunk):
+            batch.set_batch(
+                chunk.tokens,
+                n_past,
+                logits_all=False,
+                logits_last=is_last_chunk
+            )
+            ctx.decode(batch)
+
+            n_past += chunk.n_tokens
+        else:
+            new_n_past = llama_cpp.llama_pos(0)
+            result = mtmd.mtmd_helper_eval_chunk_single(
+                mctx,
+                ctx.ctx,
+                chunk.chunk_ptr,
+                llama_cpp.llama_pos(n_past),
+                llama_cpp.llama_seq_id(0),
+                n_batch,
+                False, # logits_last
+                ctypes.byref(new_n_past)
+            )
+            if result != 0:
+                raise RuntimeError(f"MTMD eval error: {result}")
+
+            n_past = new_n_past.value
