nexaai 1.0.17rc10__cp310-cp310-macosx_14_0_universal2.whl → 1.0.18__cp310-cp310-macosx_14_0_universal2.whl

nexaai/_version.py

@@ -1,4 +1,4 @@
 # This file is generated by CMake from _version.py.in
 # Do not modify this file manually - it will be overwritten
 
-__version__ = "1.0.17-rc10" 
+__version__ = "1.0.18" 

nexaai/mlx_backend/vlm/generate_qwen3_vl.py

@@ -41,7 +41,12 @@ except ImportError:
 from ml import ChatMessage
 from dataclasses import dataclass
 from typing import Any, Generator, List, Optional, Sequence, Tuple, Union
-from .generate import GenerationResult 
+from .generate import GenerationResult
+
+# Custom exception for context length exceeded
+class ContextLengthExceededError(Exception):
+    """Raised when input context length exceeds model's maximum context size"""
+    pass 
 
 @dataclass
 class Qwen3VLBundledModel:
@@ -67,6 +72,7 @@ def load_qwen3_vl(
 
     Parameters are aligned with .generate.load for compatibility.
     """
+    
     model_path = Path(path_or_repo)
     if not model_path.exists():
         if "/" in path_or_repo:
@@ -154,7 +160,6 @@ def load_qwen3_vl(
     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)
 
@@ -166,11 +171,15 @@ def load_qwen3_vl(
 
 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:
+    for i, msg in enumerate(messages):
         content_items = [{"type": "text", "text": msg.content}]
         messages_dict.append({"role": msg.role, "content": content_items})
-    return json.dumps(messages_dict)
+    
+    result = json.dumps(messages_dict)
+    
+    return result
 
 
 def stream_generate_qwen3_vl(
@@ -184,15 +193,22 @@ def stream_generate_qwen3_vl(
 
 ) -> Generator[Any, None, None]:
     """Stream generation yielding .generate.GenerationResult-compatible chunks."""
-    messages = json.loads(prompt)
+    
+    try:
+        messages = json.loads(prompt)
+    except json.JSONDecodeError as e:
+        raise
+    
     if image is not None:
         image_list = image if isinstance(image, list) else [image]
         pil_images = []
-        for p in image_list:
+        for i, p in enumerate(image_list):
             try:
-                pil_images.append(Image.open(p))
-            except Exception:
+                img = Image.open(p)
+                pil_images.append(img)
+            except Exception as e:
                 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):
@@ -201,6 +217,7 @@ def stream_generate_qwen3_vl(
 
     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")
@@ -208,10 +225,18 @@ def stream_generate_qwen3_vl(
     input_ids = inputs["input_ids"]
     pixel_values = inputs.get("pixel_values")
     image_grid_thw = inputs.get("image_grid_thw")
+    
+    
+    # Check if input context exceeds KV cache size and raise error
+    max_kv_size = 4096  # This should match the max_kv_size used in make_prompt_cache and nexa_generate_step
+    if input_ids.size > max_kv_size:
+        error_msg = f"Input context length ({input_ids.size} tokens) exceeds maximum supported context size ({max_kv_size} tokens). Please reduce the input length."
+        raise ContextLengthExceededError(error_msg)
 
     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
@@ -222,37 +247,45 @@ def stream_generate_qwen3_vl(
 
     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),
-        )
+    try:
+        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
+
+            current_tps = gen_count / max(1e-6, (time.perf_counter() - tic))
+            
+            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(current_tps),
+                peak_memory=float(mx.get_peak_memory() / 1e9),
+            )
+    except Exception as e:
+        import traceback
+        traceback.print_exc()
+        raise
+        
 
 def quant_predicate(path: str, mod: nn.Module) -> bool:
     """Quantization predicate to exclude certain layers from quantization."""

nexaai/mlx_backend/vlm/interface.py

@@ -25,7 +25,7 @@ from profiling import ProfilingMixin, ProfilingData, StopReason
 
 # Import from the actual mlx_vlm structure
 from .generate import generate, stream_generate, load
-from .generate_qwen3_vl import apply_chat_template_qwen3_vl, stream_generate_qwen3_vl, load_qwen3_vl
+from .generate_qwen3_vl import apply_chat_template_qwen3_vl, stream_generate_qwen3_vl, load_qwen3_vl, ContextLengthExceededError
 
 from .modeling.prompt_utils import apply_chat_template
 
@@ -124,7 +124,7 @@ class VLM(ProfilingMixin):
         prompt: str,
         config: Optional[GenerationConfig] = None,
     ) -> GenerationResult:
-        """Generate text from prompt."""
+        """Generate text from prompt.""" 
         # Start profiling
         self._start_profiling()
 
@@ -148,12 +148,15 @@ class VLM(ProfilingMixin):
         # Extract incremental portion of the prompt (similar to llama.cpp VLM)
         full_prompt_len = len(prompt)
         incremental_prompt = prompt
-        
-        if self.global_n_past_chars < full_prompt_len:
-            incremental_prompt = prompt[self.global_n_past_chars:]
-        else:
-            # No new text to process
-            incremental_prompt = ""
+          
+        # Apply incremental processing only for non-qwen3vl models
+        # qwen3vl requires complete JSON conversation structure
+        if self.model_name != "qwen3vl":
+            if self.global_n_past_chars < full_prompt_len:
+                incremental_prompt = prompt[self.global_n_past_chars:]
+            else:
+                # No new text to process
+                incremental_prompt = ""
 
         # End prompt processing, start decode
         self._prompt_end()
@@ -196,13 +199,15 @@ class VLM(ProfilingMixin):
             self._update_generated_tokens(generated_tokens)
             self._set_stop_reason(StopReason.ML_STOP_REASON_COMPLETED)
             
-            # Update global character position
-            self.global_n_past_chars = full_prompt_len + len(text)
+            # Update global character position (not needed for qwen3vl JSON processing)
+            if self.model_name != "qwen3vl":
+                old_pos = self.global_n_past_chars
+                self.global_n_past_chars = full_prompt_len + len(text)
             
             self._decode_end()
             self._end_profiling()
 
-            return GenerationResult(
+            result = GenerationResult(
                 text=text,
                 prompt_tokens=prompt_tokens,
                 generation_tokens=generated_tokens,
@@ -211,7 +216,18 @@ class VLM(ProfilingMixin):
                 generation_tps=stats.get("generation_tps", 0.0),
                 peak_memory=stats.get("peak_memory", 0.0),
             )
+            
+            return result
+            
+        except ContextLengthExceededError as e:
+            self._set_stop_reason(StopReason.ML_STOP_REASON_UNKNOWN)
+            self._decode_end()
+            self._end_profiling()
+            # Re-raise the original exception without wrapping it
+            raise e
         except Exception as e:
+            import traceback
+            traceback.print_exc()
             self._set_stop_reason(StopReason.ML_STOP_REASON_UNKNOWN)
             self._decode_end()
             self._end_profiling()
@@ -224,6 +240,7 @@ class VLM(ProfilingMixin):
         on_token: Optional[TokenCallback],
     ) -> GenerationResult:
         """Generate text with streaming callback. Unified method for both text and multimodal generation."""
+        
         # Start profiling
         self._start_profiling()
 
@@ -236,6 +253,7 @@ class VLM(ProfilingMixin):
         if self.sampler_config is not None:
             gen_kwargs.update(self.sampler_config.__dict__)
 
+
         # Get image and audio paths from config
         image_paths = config.image_paths if config else None
         audio_paths = config.audio_paths if config else None
@@ -244,15 +262,20 @@ class VLM(ProfilingMixin):
         image_list = [str(path) for path in image_paths] if image_paths else None
         audio_list = [str(path) for path in audio_paths] if audio_paths else None
 
+
         # Extract incremental portion of the prompt (similar to llama.cpp VLM)
         full_prompt_len = len(prompt)
         incremental_prompt = prompt
         
-        if self.global_n_past_chars < full_prompt_len:
-            incremental_prompt = prompt[self.global_n_past_chars:]
-        else:
-            # No new text to process
-            incremental_prompt = ""
+        
+        # Apply incremental processing only for non-qwen3vl models
+        # qwen3vl requires complete JSON conversation structure
+        if self.model_name != "qwen3vl":
+            if self.global_n_past_chars < full_prompt_len:
+                incremental_prompt = prompt[self.global_n_past_chars:]
+            else:
+                # No new text to process
+                incremental_prompt = ""
 
         # End prompt processing, start decode
         self._prompt_end()
@@ -264,6 +287,8 @@ class VLM(ProfilingMixin):
         stream_generate_impl = stream_generate_qwen3_vl if self.model_name == "qwen3vl" else stream_generate
 
         try:
+            token_count = 0
+            
             for result in stream_generate_impl(
                 self.model,
                 self.processor,
@@ -272,7 +297,9 @@ class VLM(ProfilingMixin):
                 audio=audio_list,
                 **gen_kwargs,
             ):
-                                # Record TTFT on first token
+                token_count += 1
+                
+                # Record TTFT on first token
                 if first_token:
                     self._record_ttft()
                     first_token = False
@@ -285,6 +312,7 @@ class VLM(ProfilingMixin):
                 text += result.text
                 last_result = result
 
+
             # Set stop reason if not user stop
             if self._profiling_context.stop_reason != StopReason.ML_STOP_REASON_USER:
                 self._set_stop_reason(StopReason.ML_STOP_REASON_EOS)
@@ -294,13 +322,15 @@ class VLM(ProfilingMixin):
                 self._update_prompt_tokens(last_result.prompt_tokens)
                 self._update_generated_tokens(last_result.generation_tokens)
 
-            # Update global character position
-            self.global_n_past_chars = full_prompt_len + len(text)
+            # Update global character position (not needed for qwen3vl JSON processing)
+            if self.model_name != "qwen3vl":
+                old_pos = self.global_n_past_chars
+                self.global_n_past_chars = full_prompt_len + len(text)
 
             self._decode_end()
             self._end_profiling()
 
-            return GenerationResult(
+            result = GenerationResult(
                 text=text,
                 token=last_result.token if last_result else None,
                 logprobs=last_result.logprobs if last_result else None,
@@ -311,7 +341,18 @@ class VLM(ProfilingMixin):
                 generation_tps=last_result.generation_tps if last_result else 0.0,
                 peak_memory=last_result.peak_memory if last_result else 0.0,
             )
+            
+            return result
+            
+        except ContextLengthExceededError as e:
+            self._set_stop_reason(StopReason.ML_STOP_REASON_UNKNOWN)
+            self._decode_end()
+            self._end_profiling()
+            # Re-raise the original exception without wrapping it
+            raise e
         except Exception as e:
+            import traceback
+            traceback.print_exc()
             self._set_stop_reason(StopReason.ML_STOP_REASON_UNKNOWN)
             self._decode_end()
             self._end_profiling()

{nexaai-1.0.17rc10.dist-info → nexaai-1.0.18.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.17rc10
+Version: 1.0.18
 Summary: Python bindings for NexaSDK C-lib backend
 Author-email: "Nexa AI, Inc." <dev@nexa.ai>
 Project-URL: Homepage, https://github.com/NexaAI/nexasdk-bridge

{nexaai-1.0.17rc10.dist-info → nexaai-1.0.18.dist-info}/RECORD

@@ -1,6 +1,6 @@
 nexaai/__init__.py,sha256=L8oB7GFZZMGnUpCg0PecDbI_ycKuQak-ZEJ4Y12_QIw,2184
-nexaai/_stub.cpython-310-darwin.so,sha256=jjti0xHLcO3ae_-qdL8tCrb2fuoMXrSSgWl6eSfh-ao,66768
-nexaai/_version.py,sha256=nKUy6Z6ytRT-zOfWuQqpmzK7vKqdC96nmMqwtCJBwMM,144
+nexaai/_stub.cpython-310-darwin.so,sha256=ngIsqeIGg05XjU7iujJRsdP5zlEM5CsU9pVEh5LBDxY,66768
+nexaai/_version.py,sha256=u4x0epv_LKPUfQvNf5zaekZDwaMD-RVDvPvcAmx-b40,139
 nexaai/asr.py,sha256=NljMXDErwPNMOPaRkJZMEDka9Nk8xyur7L8i924TStY,2054
 nexaai/base.py,sha256=N8PRgDFA-XPku2vWnQIofQ7ipz3pPlO6f8YZGnuhquE,982
 nexaai/common.py,sha256=Y0NJNLTi4Nq4x1WL6PQsSvGUto0eGmWhjpsC6jcekfA,3444
@@ -19,7 +19,7 @@ nexaai/asr_impl/pybind_asr_impl.py,sha256=pE9Hb_hMi5yAc4MF83bLVOb8zDtreCkB3_u7XE
 nexaai/binds/__init__.py,sha256=eYuay_8DDXeOUWz2_R9HFSabohxs6hvZn391t2L0Po0,104
 nexaai/binds/common_bind.cpython-310-darwin.so,sha256=zxJuD0nSV--VZKxBfWZUavU7_bHj_JTi0FhkjvG4VJw,235264
 nexaai/binds/embedder_bind.cpython-310-darwin.so,sha256=tPa0c0Dv_GiW66fgmAGWGCHXRGNApznqoQS0eQx9GFM,202064
-nexaai/binds/libnexa_bridge.dylib,sha256=ny1RbyxrSMK20bDgv6wSEwPzt5r_5bBc3uruuMH9Ufg,251192
+nexaai/binds/libnexa_bridge.dylib,sha256=uxHcfOFvu1TdqJ5GpzffLshS2-ft5EOfO6FIqthTU5M,251192
 nexaai/binds/llm_bind.cpython-310-darwin.so,sha256=TAWfa1Hzq00TjtC1xVsiAeLp6hv2LrL5afDz4omUghc,182784
 nexaai/binds/vlm_bind.cpython-310-darwin.so,sha256=nd6eG_m2EiPthzkSZ97hlXWUOZQir4cQfFJZ4p6eR2U,182704
 nexaai/binds/nexa_llama_cpp/libggml-base.dylib,sha256=JM4oOkie1su0ES5hMdtILeQHlRukRzH1vTleTupUXhg,650736
@@ -29,7 +29,7 @@ nexaai/binds/nexa_llama_cpp/libggml.dylib,sha256=aOTj_6RrAMkfDO0ZI28_3nfcC-l4Y3d
 nexaai/binds/nexa_llama_cpp/libllama.dylib,sha256=RkBd5usb8RvEIOamvxCW3UvMauI5bC66G_n6hw83NpY,1786128
 nexaai/binds/nexa_llama_cpp/libmtmd.dylib,sha256=o6mQqefzQNF0CS4j6odwJKj0gkXm15hIxwlNt88FOn4,605248
 nexaai/binds/nexa_llama_cpp/libnexa_plugin.dylib,sha256=f-ctmVHSL5x1LKLr0StYM5WscvUW_5olJ7al3ciVmmQ,1863000
-nexaai/binds/nexa_mlx/libnexa_plugin.dylib,sha256=RwkU2ZKm6T0Y6btBD7NaDM67A-MifPkkCgO9VFcpUFY,642456
+nexaai/binds/nexa_mlx/libnexa_plugin.dylib,sha256=KKM4j2JmHQrCQ1LyqEDRxwADRXKi_HM0IY5I5KFl2e0,659288
 nexaai/binds/nexa_mlx/py-lib/ml.py,sha256=DKXVOAfh8cg7KTKljh7jpcPwfQFNigc6uv_ZXF6lse8,23977
 nexaai/binds/nexa_mlx/py-lib/profiling.py,sha256=Dc-mybFwBdCIKFWL7CbSHjkOJGAoYHG7r_e_XPhzwBU,9361
 nexaai/binds/nexa_mlx/py-lib/mlx_audio/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -192,7 +192,7 @@ nexaai/binds/nexa_nexaml/libmp3lame.0.dylib,sha256=G-21u3MAZ2hiv1fbPEpOUGdToZfLt
 nexaai/binds/nexa_nexaml/libmpg123.0.dylib,sha256=L4AgLcdcjGOQkEovUGDCNlH959500i0GkKBhqiHxBY4,306464
 nexaai/binds/nexa_nexaml/libnexa-mm-process.dylib,sha256=a_63nw3Fmdlw1KoxsfwS76ZXZgqQhw4sQ8rE42hpBwo,10205704
 nexaai/binds/nexa_nexaml/libnexa-sampling.dylib,sha256=OliTBr-r2y22Ebjc1Y33Ax56XpLMFQKtsHgfouj6TQM,7865400
-nexaai/binds/nexa_nexaml/libnexa_plugin.dylib,sha256=41VEU8yhk2aQlpfoSFeYkhLLhBSbzCcMns0IGyZvpNg,249848
+nexaai/binds/nexa_nexaml/libnexa_plugin.dylib,sha256=LpNEwlFTaAVyFi2l68X_uDCO-Sr5y94obKZpPucYku4,249848
 nexaai/binds/nexa_nexaml/libnexaproc.dylib,sha256=Pnxgm2k29hfedLjNP4t44gvioh-NWbE_3BW-VsCdHp8,880560
 nexaai/binds/nexa_nexaml/libomp.dylib,sha256=2chMvsFutcZX_W9odGciLJDz2vZ2d75i6v5562I9Fqs,754912
 nexaai/binds/nexa_nexaml/libqwen3-vl.dylib,sha256=hd8VPINYTH8mbU3BPxnjrPOKTw-NkWn7zkDB7p074-U,753168
@@ -414,8 +414,8 @@ nexaai/mlx_backend/tts/__init__.py,sha256=fuT_9_xpYJ28m4yjly5L2jChUrzlSQz-b_S7nu
 nexaai/mlx_backend/tts/interface.py,sha256=0FvZbIyOvg8jERZEQ6bygbv7v02O9xHO4-TPUlar0b4,9568
 nexaai/mlx_backend/vlm/__init__.py,sha256=_25kvMEviX16Hg3bro8Ws70V0eeIEqYKV8ZDXqYzKew,73
 nexaai/mlx_backend/vlm/generate.py,sha256=DqHFEAuqk-nko8ho6U9GAXTDAWz4d8GTe_hCt-XFyCw,19071
-nexaai/mlx_backend/vlm/generate_qwen3_vl.py,sha256=undjso1mfxqpd6FMTksSA5qagRttxAGbOBj1x7cqI1s,9211
-nexaai/mlx_backend/vlm/interface.py,sha256=0BLfodbYOU71jFvAvv01FuLBE_KBtyB-8Cd7LqzzRHY,17450
+nexaai/mlx_backend/vlm/generate_qwen3_vl.py,sha256=eeizW18u6dHPZOOnJtQUJkiqMAIIpOSS-IOjacXGsz4,10240
+nexaai/mlx_backend/vlm/interface.py,sha256=HOPzWNMs6QaHO6x0Z83kW1xkRRmb8_xo6xQLKsOWqAo,19013
 nexaai/mlx_backend/vlm/main.py,sha256=nPcg25jupeDD74uvRoxpWp3Dsulw7WddI7vll6zejak,10664
 nexaai/mlx_backend/vlm/modeling/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/mlx_backend/vlm/modeling/convert.py,sha256=ia5i9cgTufFGmKyhkYUaW0nfNqT_bMo8i-Hg_zy5JC4,1863
@@ -555,7 +555,7 @@ nexaai/utils/quantization_utils.py,sha256=FYcNSAKGlBqFDUTx3jSKOr2lnq4nyiyC0ZG8oS
 nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/vlm_impl/mlx_vlm_impl.py,sha256=pLtWm_ckz8a0U-AtAOMVseFDO4OVPvHyYO2KlfBaGYk,10833
 nexaai/vlm_impl/pybind_vlm_impl.py,sha256=FAbhpRJzHgI78r0mUvKybO97R1szvNhH0aTn_I52oT4,8597
-nexaai-1.0.17rc10.dist-info/METADATA,sha256=CBp42bC2oj1pRu7t-v7qqiH6ZlQ1QLFHSpI3QL1JypU,1202
-nexaai-1.0.17rc10.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
-nexaai-1.0.17rc10.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.17rc10.dist-info/RECORD,,
+nexaai-1.0.18.dist-info/METADATA,sha256=yh4CFZmHv1dg0aN41La3qaRlXU5XC1_7erEOspSE95s,1198
+nexaai-1.0.18.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
+nexaai-1.0.18.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.18.dist-info/RECORD,,