nexaai 1.0.16rc12__cp310-cp310-macosx_14_0_universal2.whl → 1.0.16rc14__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__ = "1.0.16-rc12" 
+__version__ = "1.0.16-rc14" 

nexaai/common.py

@@ -7,6 +7,7 @@ class PluginID(str, Enum):
     """Enum for plugin identifiers."""
     MLX = "mlx"
     LLAMA_CPP = "llama_cpp"
+    NEXAML = "nexaml"
 
 
 class ChatMessage(TypedDict):

nexaai/mlx_backend/vlm/interface.py

@@ -80,6 +80,9 @@ class VLM(ProfilingMixin):
 
         # Init deafutl sampler config with defualt.
         self.sampler_config = SamplerConfig()
+        
+        # Track global character position for incremental processing
+        self.global_n_past_chars = 0
 
     def destroy(self) -> None:
         """Destroy the model and free resources."""
@@ -89,6 +92,7 @@ class VLM(ProfilingMixin):
     def reset(self) -> None:
         """Reset the model state."""
         self._reset_cache()
+        self.global_n_past_chars = 0
 
     def _reset_cache(self) -> None:
         """Reset the KV cache."""
@@ -141,6 +145,16 @@ 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 = ""
+
         # End prompt processing, start decode
         self._prompt_end()
         self._decode_start()
@@ -152,7 +166,7 @@ class VLM(ProfilingMixin):
             text, stats = generate(
                 self.model,
                 self.processor,
-                prompt,
+                incremental_prompt,  # Use incremental prompt instead of full prompt
                 image=image_list,
                 audio=audio_list,
                 **gen_kwargs,
@@ -181,6 +195,10 @@ class VLM(ProfilingMixin):
             self._update_prompt_tokens(prompt_tokens)
             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)
+            
             self._decode_end()
             self._end_profiling()
 
@@ -226,6 +244,16 @@ 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 = ""
+
         # End prompt processing, start decode
         self._prompt_end()
         self._decode_start()
@@ -239,7 +267,7 @@ class VLM(ProfilingMixin):
             for result in stream_generate_impl(
                 self.model,
                 self.processor,
-                prompt,
+                incremental_prompt,  # Use incremental prompt instead of full prompt
                 image=image_list,
                 audio=audio_list,
                 **gen_kwargs,
@@ -266,6 +294,9 @@ 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)
+
             self._decode_end()
             self._end_profiling()
 

nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/generate.py

@@ -232,7 +232,7 @@ def generate_step(
             prompt_progress_callback(prompt_processed_tokens, total_prompt_tokens)
             prompt_processed_tokens += prefill_step_size
             y = y[prefill_step_size:]
-            mx.metal.clear_cache()
+            mx.clear_cache()
 
         y, logprobs = _step(y)
 
@@ -249,7 +249,7 @@ def generate_step(
             break
         yield y.item(), logprobs
         if n % 256 == 0:
-            mx.metal.clear_cache()
+            mx.clear_cache()
         y, logprobs = next_y, next_logprobs
         n += 1
 
@@ -371,7 +371,7 @@ def nexa_generate_step(
             prompt_progress_callback(prompt_processed_tokens, total_prompt_tokens)
             prompt_processed_tokens += prefill_step_size
             y = y[prefill_step_size:]
-            mx.metal.clear_cache()
+            mx.clear_cache()
 
         y, logprobs = _step(y)
 
@@ -388,7 +388,7 @@ def nexa_generate_step(
             break
         yield y.item(), logprobs
         if n % 256 == 0:
-            mx.metal.clear_cache()
+            mx.clear_cache()
         y, logprobs = next_y, next_logprobs
         n += 1
 
@@ -507,7 +507,7 @@ def nexa_multimodal_generate_step(
             prompt_progress_callback(prompt_processed_tokens, total_prompt_tokens)
             prompt_processed_tokens += prefill_step_size
             y = y[prefill_step_size:]
-            mx.metal.clear_cache()
+            mx.clear_cache()
 
         y, logprobs = _step(y)
 
@@ -524,7 +524,7 @@ def nexa_multimodal_generate_step(
             break
         yield y.item(), logprobs
         if n % 256 == 0:
-            mx.metal.clear_cache()
+            mx.clear_cache()
         y, logprobs = next_y, next_logprobs
         n += 1
 
@@ -632,7 +632,7 @@ def speculative_generate_step(
             quantize_cache_fn(cache)
             mx.eval([c.state for c in cache])
             y = y[prefill_step_size:]
-            mx.metal.clear_cache()
+            mx.clear_cache()
         return y
 
     def _rewind_cache(num_draft, num_accept):

nexaai/utils/quantization_utils.py

@@ -25,12 +25,15 @@ class QuantizationType(str, Enum):
     F16 = "F16"
     Q2_K = "Q2_K"
     Q2_K_L = "Q2_K_L"
+    Q3_K = "Q3_K"
     Q3_K_M = "Q3_K_M"
     Q3_K_S = "Q3_K_S"
     Q4_0 = "Q4_0"
     Q4_1 = "Q4_1"
+    Q4_K = "Q4_K"
     Q4_K_M = "Q4_K_M"
     Q4_K_S = "Q4_K_S"
+    Q5_K = "Q5_K"
     Q5_K_M = "Q5_K_M"
     Q5_K_S = "Q5_K_S"
     Q6_K = "Q6_K"
@@ -67,12 +70,15 @@ def extract_quantization_from_filename(filename: str) -> Optional[QuantizationTy
         'f16.': QuantizationType.F16,  # Add F16 support
         'q2_k_l.': QuantizationType.Q2_K_L,  # Check Q2_K_L before Q2_K to avoid partial match
         'q2_k.': QuantizationType.Q2_K,
+        'q3_k.': QuantizationType.Q3_K,
         'q3_k_m.': QuantizationType.Q3_K_M,
-        'q3_ks.': QuantizationType.Q3_K_S,
+        'q3_k_s.': QuantizationType.Q3_K_S,
         'q4_k_m.': QuantizationType.Q4_K_M,
         'q4_k_s.': QuantizationType.Q4_K_S,
         'q4_0.': QuantizationType.Q4_0,
         'q4_1.': QuantizationType.Q4_1,
+        'q4_k.': QuantizationType.Q4_K,
+        'q5_k.': QuantizationType.Q5_K,
         'q5_k_m.': QuantizationType.Q5_K_M,
         'q5_k_s.': QuantizationType.Q5_K_S,
         'q6_k.': QuantizationType.Q6_K,

{nexaai-1.0.16rc12.dist-info → nexaai-1.0.16rc14.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.16rc12
+Version: 1.0.16rc14
 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-1.0.16rc12.dist-info → nexaai-1.0.16rc14.dist-info}/RECORD

@@ -1,9 +1,9 @@
 nexaai/__init__.py,sha256=L8oB7GFZZMGnUpCg0PecDbI_ycKuQak-ZEJ4Y12_QIw,2184
-nexaai/_stub.cpython-310-darwin.so,sha256=bzuFvUH92pOTqDVeEN9Y9ULBe5iQhE5RUgzD_MRDrtE,66768
-nexaai/_version.py,sha256=VcEloSyZAF14s_ZLUSzu8vLdcbhwdSK1deBNqQktuqA,144
+nexaai/_stub.cpython-310-darwin.so,sha256=8r5dERkhD_MbKv5kOuGSgx_lGZyRFZ3MxOEOdFb2hg8,66768
+nexaai/_version.py,sha256=bFprtDX2rUyZwPnP3h4-J3LGlRF9uZAd4KOCazdh12I,144
 nexaai/asr.py,sha256=NljMXDErwPNMOPaRkJZMEDka9Nk8xyur7L8i924TStY,2054
 nexaai/base.py,sha256=N8PRgDFA-XPku2vWnQIofQ7ipz3pPlO6f8YZGnuhquE,982
-nexaai/common.py,sha256=yBnIbqYaQYnfrl7IczOBh6MDibYZVxwaRJEglYcKgGs,3422
+nexaai/common.py,sha256=Y0NJNLTi4Nq4x1WL6PQsSvGUto0eGmWhjpsC6jcekfA,3444
 nexaai/cv.py,sha256=RHCDo8gvBH8BkGZx7qVyp-OKxqi7E1GG9XzyaXehCNA,3273
 nexaai/embedder.py,sha256=Cw0tSHkPgd-RI62afCqQAcTHMnQhaI2CvfTMO-1JKOg,2452
 nexaai/image_gen.py,sha256=0C_5Tjj4BYmxLbmMmvwajp-yy2mmEEOKwBFnDQNPzx4,4356
@@ -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=ya8gz7NnUtHlniGnRyExdwWfdHimEHiJ7Dry4I7_y44,235264
 nexaai/binds/embedder_bind.cpython-310-darwin.so,sha256=tPa0c0Dv_GiW66fgmAGWGCHXRGNApznqoQS0eQx9GFM,202064
-nexaai/binds/libnexa_bridge.dylib,sha256=y0cv3WWCPmKVyZFSj81BM6-yhzUOfkZWvezu_kfztQc,251192
+nexaai/binds/libnexa_bridge.dylib,sha256=iTPtrDD83Mb8KCQr521J_ggOIJu_xPnVED7yg8jkgyg,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=s4oelR1kkFrk0MKL23I0m9Uei4yPf48rzdNvknpjodg,623192
+nexaai/binds/nexa_mlx/libnexa_plugin.dylib,sha256=RwkU2ZKm6T0Y6btBD7NaDM67A-MifPkkCgO9VFcpUFY,642456
 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
@@ -182,6 +182,16 @@ nexaai/binds/nexa_mlx/py-lib/mlx_audio/tts/tests/test_base.py,sha256=dAeEA0IsBnU
 nexaai/binds/nexa_mlx/py-lib/mlx_audio/tts/tests/test_convert.py,sha256=79ddUhtTUlElD9NvBupUxl-MV4HKFM9PUxQpTwKLnBA,5804
 nexaai/binds/nexa_mlx/py-lib/mlx_audio/tts/tests/test_interpolate.py,sha256=9dNmH03C46HtxwesH2DpT2oTNEG1KCZWYEKq6UQ3vfk,3536
 nexaai/binds/nexa_mlx/py-lib/mlx_audio/tts/tests/test_models.py,sha256=12RiOfPtSZQj5g5JM-yCJk3uGQfM3OdmRiPt5uUDE4E,35096
+nexaai/binds/nexa_nexaml/libggml-base.dylib,sha256=JM4oOkie1su0ES5hMdtILeQHlRukRzH1vTleTupUXhg,650736
+nexaai/binds/nexa_nexaml/libggml-cpu.so,sha256=qiYxbTe4Nt7n36zJVvq3zovgSZEmrN2is6gzTern7UI,677728
+nexaai/binds/nexa_nexaml/libggml-metal.so,sha256=zfaX7rIBYQazH2lf-vza007BMhPTK1ASd2T0HLLIA4E,673104
+nexaai/binds/nexa_nexaml/libggml.dylib,sha256=aOTj_6RrAMkfDO0ZI28_3nfcC-l4Y3dRCiS3C0d0_eI,58592
+nexaai/binds/nexa_nexaml/libnexa-mm-process.dylib,sha256=3mITty3oYhoi5yc690CxS0tf4LGanqrPCR0meE-h9HQ,8568120
+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/libnexaproc.dylib,sha256=Pnxgm2k29hfedLjNP4t44gvioh-NWbE_3BW-VsCdHp8,880560
+nexaai/binds/nexa_nexaml/libqwen3-vl.dylib,sha256=hd8VPINYTH8mbU3BPxnjrPOKTw-NkWn7zkDB7p074-U,753168
+nexaai/binds/nexa_nexaml/libqwen3vl-vision.dylib,sha256=AmUk-DBtBAp5DOR14ICZcsiAcKjxxAx1y6mHAaETQu8,570024
 nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/cv_impl/mlx_cv_impl.py,sha256=gKECQOv8iaWwG3bl7xeqVy2NN_9K7tYerIFzfn4eLo4,3228
 nexaai/cv_impl/pybind_cv_impl.py,sha256=uSmwBste4cT7c8DQmXzRLmzwDf773PAbXNYWW1UzVls,1064
@@ -400,7 +410,7 @@ nexaai/mlx_backend/tts/interface.py,sha256=0FvZbIyOvg8jERZEQ6bygbv7v02O9xHO4-TPU
 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=vFTzJCbqq55ybv_tbDBC9NVn1_sXgCfqXdsV-3ia8vo,16177
+nexaai/mlx_backend/vlm/interface.py,sha256=0BLfodbYOU71jFvAvv01FuLBE_KBtyB-8Cd7LqzzRHY,17450
 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
@@ -515,7 +525,7 @@ nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/qwen3vl.py,sha256=LArnNtI98B_GJO
 nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/base.py,sha256=4RlZwgz8YX2ngmJNaymxFFpw9hJu-0EMw9xwXpngW9o,3496
 nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/cache.py,sha256=NMOB6x-RT6svF4H-Ymo5WqnP7ptAal3aaKjWZXWGMsM,17671
-nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/generate.py,sha256=Mw7Btz0_t7erQOrfWzCXT-ktEwZl61OODcmDMIo3VS0,26719
+nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/generate.py,sha256=bchCpnlewysWQss5TQKxdKPXYd5VA7ySUDfRt8Xj_H4,26677
 nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/rope_utils.py,sha256=ty0dA3SsEUFtFbHo16tKdnKymrNKKsUO3KMYapMajbY,8704
 nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/sample_utils.py,sha256=8SEeVwgjuvaYy-4ALAU0RHQMuRr2k7EkXba_csxk498,10673
 nexaai/mlx_backend/vlm/modeling/models/qwen3_vl/llm_common/tokenizer_utils.py,sha256=Gqanx4hBDcon_k5ClhUsS4YpMbZNiee8jvImGS9h43s,13229
@@ -537,11 +547,11 @@ nexaai/utils/manifest_utils.py,sha256=sR9Nme4GbD3Cb3fMd55yLvGZpqxb71vd6b2XZTsrIG
 nexaai/utils/model_manager.py,sha256=p2kJKK63Zk-rEUucFsgY0T5PyXi_IvJY0gKewUVcAV4,56081
 nexaai/utils/model_types.py,sha256=-DER8L4lAUR_iLS99F0r57avwqWtuN21ug5pX2p24_E,1369
 nexaai/utils/progress_tracker.py,sha256=jdUqtmPqyhwC9uSKvQcJEYETwSt-OhP4oitdJ94614o,15394
-nexaai/utils/quantization_utils.py,sha256=4gvp6UQfSO9G1FYBwnFtQspTzH9sDbi1PBXw2t1N69M,7650
+nexaai/utils/quantization_utils.py,sha256=FYcNSAKGlBqFDUTx3jSKOr2lnq4nyiyC0ZG8oSxFwiU,7825
 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-1.0.16rc12.dist-info/METADATA,sha256=a6Oh67bHm9t2m1j9yPHyou4gx0YOJjag7YLhJYyjRxc,1202
-nexaai-1.0.16rc12.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
-nexaai-1.0.16rc12.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.16rc12.dist-info/RECORD,,
+nexaai-1.0.16rc14.dist-info/METADATA,sha256=rD9zD2HduPUSrlSkUZUQ4Ut2g6nvwQ-PN0kgODO7TEU,1202
+nexaai-1.0.16rc14.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
+nexaai-1.0.16rc14.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.16rc14.dist-info/RECORD,,