nexaai 1.0.18rc1__cp310-cp310-macosx_14_0_universal2.whl → 1.0.19__cp310-cp310-macosx_14_0_universal2.whl

nexaai/mlx_backend/vlm/generate_qwen3_vl_moe.py

@@ -0,0 +1,259 @@
+import argparse
+import json
+import os
+import mlx.core as mx
+import mlx.nn as nn
+import time
+from PIL import Image
+import requests
+import numpy as np
+from pathlib import Path
+from huggingface_hub import snapshot_download
+from dataclasses import dataclass
+from typing import Any, Generator, List, Optional, Sequence, Tuple, Union
+
+# Import required modules for quantized loading
+from transformers import AutoTokenizer
+
+# Import from the nested modeling structure
+from .modeling.models.qwen3vl_moe.llm_common.generate import nexa_generate_step
+from .modeling.models.qwen3vl_moe.llm_common.cache import make_prompt_cache
+from .modeling.models.qwen3vl_moe.qwen3vl_moe import (
+    VEGModel, LLMModel, ModelArgs, VisionConfig, TextConfig, handle_multimodal_embeds
+)
+from .modeling.models.qwen3vl_moe.processor import Qwen3VLProcessor
+from .generate import GenerationResult
+from ml import ChatMessage 
+
+@dataclass
+class Qwen3VLBundledModel:
+    """Container for Qwen3-VL MoE vision and language models."""
+    vision_model: VEGModel
+    llm_model: LLMModel
+
+
+def _ensure_list(x: Union[str, List[str], None]) -> Optional[List[str]]:
+    if x is None:
+        return None
+    return x if isinstance(x, list) else [x]
+
+
+def load_qwen3_vl(
+    path_or_repo: str,
+    adapter_path: Optional[str] = None,
+    lazy: bool = False,
+    revision: Optional[str] = None,
+    **kwargs,
+) -> Tuple[Qwen3VLBundledModel, Qwen3VLProcessor]:
+    """Load Qwen3-VL MoE quantized models and processor.
+
+    Parameters are aligned with .generate.load for compatibility.
+    """
+    model_path = Path(path_or_repo)
+    if not model_path.exists():
+        if "/" in path_or_repo:
+            model_path = Path(snapshot_download(
+                repo_id=path_or_repo, repo_type="model", revision=revision))
+        else:
+            # Fallback to local modelfiles directory relative to this file
+            curr_dir = Path(__file__).parent
+            model_path = curr_dir / "modeling" / "models" / "qwen3vl_moe" / "modelfiles"
+            if not model_path.exists():
+                model_path = curr_dir / "modelfiles"
+
+    # Model configs - Updated to match Qwen3VL-MoE specifications
+    vision_config = VisionConfig(
+        hidden_size=1152,
+        intermediate_size=4304,
+        num_heads=16,
+        num_hidden_layers=27,
+        patch_size=16,
+        temporal_patch_size=2,
+        in_channels=3,
+        hidden_act="gelu_pytorch_tanh",
+        spatial_merge_size=2,
+        out_hidden_size=2048,
+        num_position_embeddings=2304,
+        deepstack_visual_indexes=[8, 16, 24],
+    )
+
+    text_config = TextConfig(
+        model_type="qwen3_vl_moe_text",
+        hidden_size=2048,
+        num_hidden_layers=48,
+        intermediate_size=6144,
+        num_attention_heads=32,
+        num_key_value_heads=4,
+        rms_norm_eps=1e-6,
+        vocab_size=152064,
+        max_position_embeddings=128000,
+        rope_theta=1000000.0,
+        head_dim=128,
+        tie_word_embeddings=False,
+        attention_bias=False,
+        attention_dropout=0.0,
+        rope_scaling={
+            "mrope_interleaved": True,
+            "mrope_section": [24, 20, 20],
+            "rope_type": "default"
+        },
+        # MoE specific parameters
+        num_experts=128,
+        num_experts_per_tok=8,
+        moe_intermediate_size=768,
+        shared_expert_intermediate_size=0,
+        norm_topk_prob=True,
+        decoder_sparse_step=1,
+        max_window_layers=48,
+        sliding_window=32768,
+        mlp_only_layers=[],
+        use_qk_norm=True,
+        layer_types=[],
+    )
+
+    vision_model = VEGModel(vision_config)
+    llm_model = LLMModel(text_config)
+
+    # Try to load LLM model from available files in order of preference
+    preferred_order = [
+        ("qwen3vl-moe-llm-30B-A3B-q4_0.safetensors", 4),
+        ("qwen3vl-moe-llm-30B-A3B-q8_0.safetensors", 8),
+        ("qwen3vl-moe-llm-30B-A3B-f32.safetensors", 32),
+    ]
+
+    llm_weights_path = None
+    quantization_bits = None
+    
+    # Try loading in order of preference
+    for filename, bits in preferred_order:
+        candidate_path = model_path / filename
+        if candidate_path.exists():
+            llm_weights_path = candidate_path
+            quantization_bits = bits
+            break
+    
+    if llm_weights_path is None:
+        # Fallback to original hardcoded path for backward compatibility
+        llm_weights_path = model_path / "qwen3vl-moe-llm-30B-A3B-q4_0.safetensors"
+        quantization_bits = 4
+
+    vision_weights_path = model_path / "qwen3vl-moe-vision-30B-A3B-f16.safetensors"
+
+    if not vision_weights_path.exists():
+        raise FileNotFoundError(
+            f"Missing vision weights: {vision_weights_path}"
+        )
+
+    # Load weights (vision fp16, llm with detected quantization)
+    vision_model.set_dtype(mx.float16)
+    vision_model.load_weights(str(vision_weights_path), strict=True)
+
+    # Apply quantization if needed and load LLM weights
+    if quantization_bits in [4, 8]:
+        nn.quantize(llm_model, bits=quantization_bits, group_size=64,
+                    class_predicate=quant_predicate)
+    # For f32 (32-bit), no quantization needed
+    
+    llm_model.load_weights(str(llm_weights_path), strict=True)
+
+    # Tokenizer and processor
+    tokenizer = AutoTokenizer.from_pretrained(path_or_repo)
+    processor = Qwen3VLProcessor(tokenizer=tokenizer)
+
+    return Qwen3VLBundledModel(vision_model=vision_model, llm_model=llm_model), processor
+
+def apply_chat_template_qwen3_vl(messages: Sequence[ChatMessage], num_images: int = 0, num_audios: int = 0, tools: Optional[str] = None, enable_thinking: bool = False) -> str:
+    """Apply chat template: serialize messages with content as a list of typed items."""
+    messages_dict = []
+    for msg in messages:
+        content_items = [{"type": "text", "text": msg.content}]
+        messages_dict.append({"role": msg.role, "content": content_items})
+    return json.dumps(messages_dict)
+
+
+def stream_generate_qwen3_vl(
+    model: Qwen3VLBundledModel,
+    processor: Qwen3VLProcessor,
+    prompt: str,
+    image: Union[str, List[str]] = None,
+    audio: Union[str, List[str]] = None,
+    max_tokens: int = 512,
+    **kwargs,
+
+) -> Generator[Any, None, None]:
+    """Stream generation yielding .generate.GenerationResult-compatible chunks."""
+    messages = json.loads(prompt)
+    if image is not None:
+        image_list = image if isinstance(image, list) else [image]
+        pil_images = []
+        for p in image_list:
+            try:
+                pil_images.append(Image.open(p))
+            except Exception:
+                continue
+        contents = [{"type": "image", "image": img} for img in pil_images]
+        if messages:
+            if "content" not in messages[-1] or not isinstance(messages[-1]["content"], list):
+                messages[-1]["content"] = []
+            messages[-1]["content"].extend(contents)
+
+    raw_text, processed_images = processor.messages_to_text(
+        messages, add_generation_prompt=True)
+
+    inputs = processor.text_to_input_ids(
+        raw_text, images=processed_images, return_tensors="mlx")
+
+    input_ids = inputs["input_ids"]
+    pixel_values = inputs.get("pixel_values")
+    image_grid_thw = inputs.get("image_grid_thw")
+
+    inputs_embeds, deepstack_visual_embeds, visual_pos_masks, cos, sin, rope_deltas = handle_multimodal_embeds(
+        model.vision_model, model.llm_model, input_ids, pixel_values, image_grid_thw
+    )
+
+    prompt_cache = make_prompt_cache(model.llm_model, max_kv_size=4096)
+    tokenizer = processor.tokenizer
+
+    # Rough prompt TPS estimation based on input size
+    prompt_start = time.perf_counter()
+    prompt_tps = input_ids.size / max(1e-6, (time.perf_counter() - prompt_start))
+
+    gen_count = 0
+    tic = time.perf_counter()
+
+    for token, logprobs in nexa_generate_step(
+        model=model.llm_model,
+        prompt=None,
+        input_embeddings=inputs_embeds,
+        max_tokens=max_tokens,
+        max_kv_size=4096,
+        prompt_cache=prompt_cache,
+        visual_pos_masks=visual_pos_masks,
+        deepstack_visual_embeds=deepstack_visual_embeds,
+        cos=cos,
+        sin=sin,
+        rope_deltas=rope_deltas,
+    ):
+        if token == tokenizer.eos_token_id:
+            break
+
+        text_piece = tokenizer.decode([token])
+        gen_count += 1
+
+        yield GenerationResult(
+            text=text_piece,
+            token=token,
+            logprobs=logprobs,
+            prompt_tokens=int(input_ids.size),
+            generation_tokens=gen_count,
+            prompt_tps=float(prompt_tps),
+            generation_tps=float(
+                gen_count / max(1e-6, (time.perf_counter() - tic))),
+            peak_memory=float(mx.get_peak_memory() / 1e9),
+        )
+
+def quant_predicate(path: str, mod: nn.Module) -> bool:
+    """Quantization predicate to exclude certain layers from quantization."""
+    if path.endswith("lm_head") or "norm" in path.lower() or "embed" in path.lower():
+        return False
+    return isinstance(mod, (nn.Linear, nn.Embedding))

nexaai/mlx_backend/vlm/interface.py

@@ -27,6 +27,10 @@ from profiling import ProfilingMixin, ProfilingData, StopReason
 from .generate import generate, stream_generate, load
 from .generate_qwen3_vl import apply_chat_template_qwen3_vl, stream_generate_qwen3_vl, load_qwen3_vl, ContextLengthExceededError
 
+from .generate_qwen3_vl_moe import apply_chat_template_qwen3_vl as apply_chat_template_qwen3_vl_moe
+from .generate_qwen3_vl_moe import stream_generate_qwen3_vl as stream_generate_qwen3_vl_moe
+from .generate_qwen3_vl_moe import load_qwen3_vl as load_qwen3_vl_moe
+
 from .modeling.prompt_utils import apply_chat_template
 
 # --------------------------------------------------------------------------------------
@@ -75,8 +79,15 @@ class VLM(ProfilingMixin):
         self.context_length = context_length
         self.device = device
 
-        load_impl = load_qwen3_vl if model_name == "qwen3vl" else load
-        self.model, self.processor = load_impl(str(model_path))
+        if model_name == "qwen3vl-moe":
+            load_impl = load_qwen3_vl_moe
+        elif model_name in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking"]:
+            load_impl = load_qwen3_vl
+        else:
+            load_impl = load
+
+        # Pass model_name to the loader for proper configuration
+        self.model, self.processor = load_impl(str(model_path), model_name=model_name)
 
         # Init deafutl sampler config with defualt.
         self.sampler_config = SamplerConfig()
@@ -84,6 +95,19 @@ class VLM(ProfilingMixin):
         # Track global character position for incremental processing
         self.global_n_past_chars = 0
 
+        # Add conversation state tracking to VLM class
+        if model_name in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking"]:
+            # Import here to avoid circular imports
+            from .modeling.models.qwen3_vl.llm_common.cache import make_prompt_cache
+            import mlx.core as mx
+            
+            # Initialize conversation state
+            self.rope_deltas_total = mx.zeros((1, 1), dtype=mx.int32)
+            self.prompt_cache = make_prompt_cache(self.model.llm_model, max_kv_size=4096)
+        else:
+            self.rope_deltas_total = None
+            self.prompt_cache = None
+
     def destroy(self) -> None:
         """Destroy the model and free resources."""
         self.model = None
@@ -93,6 +117,14 @@ class VLM(ProfilingMixin):
         """Reset the model state."""
         self._reset_cache()
         self.global_n_past_chars = 0
+        
+        # Reset conversation state for qwen3vl models
+        if self.model_name in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking"]:
+            import mlx.core as mx
+            from .modeling.models.qwen3_vl.llm_common.cache import make_prompt_cache
+            
+            self.rope_deltas_total = mx.zeros((1, 1), dtype=mx.int32)
+            self.prompt_cache = make_prompt_cache(self.model.llm_model, max_kv_size=4096)
 
     def _reset_cache(self) -> None:
         """Reset the KV cache."""
@@ -270,7 +302,7 @@ class VLM(ProfilingMixin):
         
         # Apply incremental processing only for non-qwen3vl models
         # qwen3vl requires complete JSON conversation structure
-        if self.model_name != "qwen3vl":
+        if self.model_name not in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking", "qwen3vl-moe"]:
             if self.global_n_past_chars < full_prompt_len:
                 incremental_prompt = prompt[self.global_n_past_chars:]
             else:
@@ -284,33 +316,70 @@ class VLM(ProfilingMixin):
         text = ""
         last_result = None
         first_token = True
-        stream_generate_impl = stream_generate_qwen3_vl if self.model_name == "qwen3vl" else stream_generate
+
+        if self.model_name == "qwen3vl-moe":
+            stream_generate_impl = stream_generate_qwen3_vl_moe
+        elif self.model_name in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking"]:
+            stream_generate_impl = stream_generate_qwen3_vl
+        else:
+            stream_generate_impl = stream_generate
 
         try:
             token_count = 0
             
-            for result in stream_generate_impl(
-                self.model,
-                self.processor,
-                incremental_prompt,  # Use incremental prompt instead of full prompt
-                image=image_list,
-                audio=audio_list,
-                **gen_kwargs,
-            ):
-                token_count += 1
-                
-                # Record TTFT on first token
-                if first_token:
-                    self._record_ttft()
-                    first_token = False
-
-                # Call the token callback if provided
-                if on_token is not None:
-                    if not on_token(result.text):
-                        self._set_stop_reason(StopReason.ML_STOP_REASON_USER)
-                        break
-                text += result.text
-                last_result = result
+            # Pass conversation state for qwen3vl models
+            if self.model_name in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking"]:
+                for result in stream_generate_impl(
+                    self.model,
+                    self.processor,
+                    incremental_prompt,
+                    image=image_list,
+                    audio=audio_list,
+                    rope_deltas_total=self.rope_deltas_total,  # Pass conversation state
+                    prompt_cache=self.prompt_cache,  # Pass KV cache
+                    **gen_kwargs,
+                ):
+                    token_count += 1
+                    
+                    # Record TTFT on first token
+                    if first_token:
+                        self._record_ttft()
+                        first_token = False
+
+                    # Call the token callback if provided
+                    if on_token is not None:
+                        if not on_token(result.text):
+                            self._set_stop_reason(StopReason.ML_STOP_REASON_USER)
+                            break
+                    text += result.text
+                    last_result = result
+
+                    # Update conversation state after each token
+                    # Note: rope_deltas_total is updated inside stream_generate_qwen3_vl
+
+            else:
+                for result in stream_generate_impl(
+                    self.model,
+                    self.processor,
+                    incremental_prompt,
+                    image=image_list,
+                    audio=audio_list,
+                    **gen_kwargs,
+                ):
+                    token_count += 1
+                    
+                    # Record TTFT on first token
+                    if first_token:
+                        self._record_ttft()
+                        first_token = False
+
+                    # Call the token callback if provided
+                    if on_token is not None:
+                        if not on_token(result.text):
+                            self._set_stop_reason(StopReason.ML_STOP_REASON_USER)
+                            break
+                    text += result.text
+                    last_result = result
 
 
             # Set stop reason if not user stop
@@ -323,7 +392,7 @@ class VLM(ProfilingMixin):
                 self._update_generated_tokens(last_result.generation_tokens)
 
             # Update global character position (not needed for qwen3vl JSON processing)
-            if self.model_name != "qwen3vl":
+            if self.model_name not in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking", "qwen3vl-moe"]:
                 old_pos = self.global_n_past_chars
                 self.global_n_past_chars = full_prompt_len + len(text)
 
@@ -428,10 +497,10 @@ class VLM(ProfilingMixin):
 
     def apply_chat_template_with_media(self, messages: Sequence[ChatMessage], num_images: int = 0, num_audios: int = 0, tools: Optional[str] = None, enable_thinking: bool = True) -> str:
         """Apply chat template to messages with proper image/audio token insertion and optional tools support."""
-        if self.model_name == "qwen3vl":
+        if self.model_name in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking"]:
             return apply_chat_template_qwen3_vl(messages, num_images=num_images, num_audios=num_audios, tools=tools, enable_thinking=enable_thinking)
-        
-        # Convert ChatMessage objects to dictionaries for the processor
+        if self.model_name == "qwen3vl-moe":
+            return apply_chat_template_qwen3_vl_moe(messages, num_images=num_images, num_audios=num_audios, tools=tools, enable_thinking=enable_thinking)
         messages_dict = [{"role": msg.role, "content": msg.content} for msg in messages]
 
         parsed_tools = None

nexaai/mlx_backend/vlm/main.py

@@ -40,6 +40,57 @@ def parse_media_from_input(user_input):
     
     return prompt, image_paths if image_paths else None, audio_paths if audio_paths else None
 
+def detect_model_name_and_repo(model_path):
+    """Detect model name and corresponding HuggingFace repo based on model path or name"""
+    model_path_lower = model_path.lower()
+    
+    # Handle HuggingFace repo format
+    if "/" in model_path:
+        repo_name = model_path.split("/")[-1] if model_path.endswith("/") else model_path.split("/")[-1]
+        repo_name_lower = repo_name.lower()
+    else:
+        repo_name_lower = model_path_lower
+    
+    # Model name mapping based on the provided examples
+    model_mappings = {
+        # 4B models
+        "qwen3vl-4b-4bit-mlx": ("qwen3vl-4b", "NexaAI/qwen3vl-4B-4bit-mlx"),
+        "qwen3vl-4b-fp16-mlx": ("qwen3vl-4b", "NexaAI/qwen3vl-4B-fp16-mlx"),
+        "qwen3vl-4b-thinking-4bit-mlx": ("qwen3vl-4b-thinking", "NexaAI/qwen3vl-4B-thinking-4bit-mlx"),
+        "qwen3vl-4b-thinking-fp16-mlx": ("qwen3vl-4b-thinking", "NexaAI/qwen3vl-4B-thinking-fp16-mlx"),
+        
+        # 8B models
+        "qwen3vl-8b-4bit-mlx": ("qwen3vl-8b", "NexaAI/qwen3vl-8B-4bit-mlx"),
+        "qwen3vl-8b-fp16-mlx": ("qwen3vl-8b", "NexaAI/qwen3vl-8B-fp16-mlx"),
+        "qwen3vl-8b-thinking-4bit-mlx": ("qwen3vl-8b-thinking", "NexaAI/qwen3vl-8B-thinking-4bit-mlx"),
+        "qwen3vl-8b-thinking-fp16-mlx": ("qwen3vl-8b-thinking", "NexaAI/qwen3vl-8B-thinking-fp16-mlx"),
+    }
+    
+    # Check exact matches first
+    for key, (model_name, repo) in model_mappings.items():
+        if key in repo_name_lower:
+            return model_name, repo if "/" not in model_path else model_path
+    
+    # Fallback detection based on patterns
+    if "qwen3vl" in repo_name_lower:
+        if "8b" in repo_name_lower:
+            if "thinking" in repo_name_lower:
+                return "qwen3vl-8b-thinking", model_path
+            else:
+                return "qwen3vl-8b", model_path
+        elif "4b" in repo_name_lower:
+            if "thinking" in repo_name_lower:
+                return "qwen3vl-4b-thinking", model_path
+            else:
+                return "qwen3vl-4b", model_path
+        else:
+            # Default to 4B if size not specified
+            return "qwen3vl-4b", model_path
+    elif "gemma" in repo_name_lower:
+        return "gemma3", model_path
+    
+    return "", model_path
+
 def parse_arguments():
     """Parse command line arguments for the VLM main function."""
     parser = argparse.ArgumentParser(
@@ -48,14 +99,14 @@ def parse_arguments():
     parser.add_argument(
         "--model_path",
         type=str,
-        default="mlx-community/gemma-3-4b-it-8bit",
+        default="NexaAI/qwen3vl-4B-4bit-mlx",
         help="The path to the local model directory or Hugging Face repo."
     )
     parser.add_argument(
         "--model_name",
         type=str,
         default="",
-        help="Specific model name/type (e.g., 'qwen3vl', 'gemma3'). If empty, auto-detect from model_path."
+        help="Specific model name/type (e.g., 'qwen3vl-4b', 'qwen3vl-4b-thinking', 'qwen3vl-8b', 'qwen3vl-8b-thinking'). If empty, auto-detect from model_path."
     )
     parser.add_argument(
         "--context_length",
@@ -89,18 +140,16 @@ def main():
     
     # Auto-detect model name if not provided
     model_name = args.model_name
+    model_path = args.model_path
+    
     if not model_name:
-        if "qwen" in args.model_path.lower():
-            model_name = "qwen3vl"
-        elif "gemma" in args.model_path.lower():
-            model_name = "gemma3"
-        else:
-            model_name = ""
+        model_name, model_path = detect_model_name_and_repo(args.model_path)
+        print(f"Auto-detected model: {model_name} from path: {model_path}")
     
     # Load the VLM instance
     vlm = VLM(
         model_name=model_name,
-        model_path=args.model_path,
+        model_path=model_path,
         mmproj_path=None,  # Not needed for this model
         context_length=args.context_length,
         device=None