lemonade-sdk 8.1.9__py3-none-any.whl → 8.1.11__py3-none-any.whl

lemonade/tools/server/wrapped_server.py

@@ -58,8 +58,22 @@ class WrappedServerTelemetry(ABC):
         telemetry = [
             ["Input tokens", self.input_tokens],
             ["Output tokens", self.output_tokens],
-            ["TTFT (s)", f"{self.time_to_first_token:.2f}"],
-            ["TPS", f"{self.tokens_per_second:.2f}"],
+            [
+                "TTFT (s)",
+                (
+                    f"{self.time_to_first_token:.2f}"
+                    if self.time_to_first_token is not None
+                    else "N/A"
+                ),
+            ],
+            [
+                "TPS",
+                (
+                    f"{self.tokens_per_second:.2f}"
+                    if self.tokens_per_second is not None
+                    else "N/A"
+                ),
+            ],
         ]
 
         table = tabulate(
@@ -83,7 +97,7 @@ class WrappedServer(ABC):
         self.telemetry: WrappedServerTelemetry = telemetry
         self.log_thread_exception = None
 
-    def choose_port(self):
+    def _choose_port(self):
         """
         Users probably don't care what port we start the wrapped server on, so let's
         search for an empty port
@@ -318,18 +332,44 @@ class WrappedServer(ABC):
         if chat_completion_request.stream:
 
             def event_stream():
-                try:
-                    # Enable streaming
-                    # pylint: disable=missing-kwoa
-                    for chunk in client.chat.completions.create(**openai_client_params):
-                        yield f"data: {chunk.model_dump_json()}\n\n"
-                    yield "data: [DONE]\n\n"
-
-                    # Show telemetry after completion
-                    self.telemetry.show_telemetry()
+                # Ensure streaming is enabled in params
+                stream_params = dict(openai_client_params)
+                stream_params["stream"] = True
 
-                except Exception as e:  # pylint: disable=broad-exception-caught
-                    yield f'data: {{"error": "{str(e)}"}}\n\n'
+                # Use streaming context so we can explicitly close on cancellation
+                with client.chat.completions.with_streaming_response.create(
+                    # pylint: disable=missing-kwoa
+                    **stream_params,
+                ) as response:
+                    try:
+                        for line in response.iter_lines():
+                            # Preserve SSE event boundaries: blank line separates events
+                            if line == b"" or line == "":
+                                yield "\n"
+                                continue
+                            if isinstance(line, bytes):
+                                try:
+                                    line = line.decode("utf-8", errors="ignore")
+                                except (UnicodeDecodeError, LookupError):
+                                    # Skip lines that fail decoding due to encoding issues
+                                    continue
+                            # Forward SSE lines as-is
+                            if not line.endswith("\n"):
+                                line += "\n"
+                            yield line
+
+                        # Show telemetry after completion
+                        self.telemetry.show_telemetry()
+
+                    except GeneratorExit:
+                        # Client disconnected/cancelled; close upstream stream and stop
+                        try:
+                            response.close()
+                        except Exception:  # pylint: disable=broad-exception-caught
+                            pass
+                        raise
+                    except Exception as e:  # pylint: disable=broad-exception-caught
+                        yield f'data: {{"error": "{str(e)}"}}\n\n'
 
             return StreamingResponse(
                 event_stream(),
@@ -387,18 +427,44 @@ class WrappedServer(ABC):
         if completion_request.stream:
 
             def event_stream():
-                try:
-                    # Enable streaming
-                    # pylint: disable=missing-kwoa
-                    for chunk in client.completions.create(**openai_client_params):
-                        yield f"data: {chunk.model_dump_json()}\n\n"
-                    yield "data: [DONE]\n\n"
-
-                    # Show telemetry after completion
-                    self.telemetry.show_telemetry()
+                # Ensure streaming is enabled in params
+                stream_params = dict(openai_client_params)
+                stream_params["stream"] = True
 
-                except Exception as e:  # pylint: disable=broad-exception-caught
-                    yield f'data: {{"error": "{str(e)}"}}\n\n'
+                # Use streaming context so we can explicitly close on cancellation
+                with client.completions.with_streaming_response.create(
+                    # pylint: disable=missing-kwoa
+                    **stream_params,
+                ) as response:
+                    try:
+                        for line in response.iter_lines():
+                            # Preserve SSE event boundaries: blank line separates events
+                            if line == b"" or line == "":
+                                yield "\n"
+                                continue
+                            if isinstance(line, bytes):
+                                try:
+                                    line = line.decode("utf-8", errors="ignore")
+                                except (UnicodeDecodeError, LookupError):
+                                    # Skip lines that fail decoding due to encoding issues
+                                    continue
+                            # Forward SSE lines as-is
+                            if not line.endswith("\n"):
+                                line += "\n"
+                            yield line
+
+                        # Show telemetry after completion
+                        self.telemetry.show_telemetry()
+
+                    except GeneratorExit:
+                        # Client disconnected/cancelled; close upstream stream and stop
+                        try:
+                            response.close()
+                        except Exception:  # pylint: disable=broad-exception-caught
+                            pass
+                        raise
+                    except Exception as e:  # pylint: disable=broad-exception-caught
+                        yield f'data: {{"error": "{str(e)}"}}\n\n'
 
             return StreamingResponse(
                 event_stream(),

lemonade/version.py

@@ -1 +1 @@
-__version__ = "8.1.9"
+__version__ = "8.1.11"

lemonade_install/install.py

@@ -447,6 +447,12 @@ class Install:
             choices=["rocm", "vulkan"],
         )
 
+        parser.add_argument(
+            "--flm",
+            action="store_true",
+            help="Install FLM (FastFlowLM) for running local language models",
+        )
+
         parser.add_argument(
             "--override",
             action="store_true",
@@ -727,19 +733,33 @@ class Install:
 
         install_llamacpp(backend)
 
+    @staticmethod
+    def _install_flm():
+        """
+        Install FLM (FastFlowLM) for running local language models.
+        """
+
+        # Check if the processor is supported before proceeding
+        check_ryzen_ai_processor()
+
+        from lemonade.tools.flm.utils import install_flm
+
+        install_flm()
+
     def run(
         self,
         ryzenai: Optional[str] = None,
         build_model: Optional[str] = None,
         llamacpp: Optional[str] = None,
+        flm: Optional[bool] = None,
         yes: bool = False,
         token: Optional[str] = None,
         override: bool = False,
     ):
-        if ryzenai is None and llamacpp is None:
+        if ryzenai is None and llamacpp is None and flm is None:
             raise ValueError(
                 "You must select something to install, "
-                "for example `--ryzenai` or `--llamacpp`"
+                "for example `--llamacpp`, `--flm`, or `--ryzenai`"
             )
 
         if ryzenai is not None:
@@ -748,6 +768,9 @@ class Install:
         if llamacpp is not None:
             self._install_llamacpp(llamacpp)
 
+        if flm:
+            self._install_flm()
+
 
 def main():
     installer = Install()

{lemonade_sdk-8.1.9.dist-info → lemonade_sdk-8.1.11.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lemonade-sdk
-Version: 8.1.9
+Version: 8.1.11
 Summary: Lemonade SDK: Your LLM Aide for Validation and Deployment
 Author-email: lemonade@amd.com
 Requires-Python: >=3.10, <3.14
@@ -29,6 +29,7 @@ Requires-Dist: tabulate
 Requires-Dist: sentencepiece
 Requires-Dist: huggingface-hub[hf_xet]==0.33.0
 Requires-Dist: python-dotenv
+Requires-Dist: rumps>=0.4.0; sys_platform == "darwin"
 Provides-Extra: oga-ryzenai
 Requires-Dist: onnxruntime-genai-directml-ryzenai==0.7.0.2.1; extra == "oga-ryzenai"
 Requires-Dist: protobuf>=6.30.1; extra == "oga-ryzenai"
@@ -65,6 +66,8 @@ Dynamic: summary
     <img src="https://img.shields.io/badge/Windows-11-0078D6?logo=windows&logoColor=white" alt="Windows 11" /></a>
   <a href="https://lemonade-server.ai/#linux" title="Ubuntu 24.04 & 25.04 Supported">
     <img src="https://img.shields.io/badge/Ubuntu-24.04%20%7C%2025.04-E95420?logo=ubuntu&logoColor=white" alt="Ubuntu 24.04 | 25.04" /></a>
+  <a href="https://lemonade-server.ai/" title="macOS 14+ with Apple Silicon">
+    <img src="https://img.shields.io/badge/macOS-14%2B-000000?logo=apple&logoColor=white" alt="macOS 14+" /></a>
   <a href="docs/README.md#installation" title="Check out our instructions">
     <img src="https://img.shields.io/badge/Python-3.10--3.13-blue?logo=python&logoColor=white" alt="Made with Python" /></a>
   <a href="https://github.com/lemonade-sdk/lemonade/blob/main/docs/contribute.md" title="Contribution Guide">
@@ -152,11 +155,11 @@ You can also import custom GGUF and ONNX models from Hugging Face by using our [
 
 Lemonade supports the following configurations, while also making it easy to switch between them at runtime. Find more information about it [here](./docs/README.md#software-and-hardware-overview).
 
-| Hardware | Engine: OGA | Engine: llamacpp | Engine: HF | Windows | Linux |
-|----------|-------------|------------------|------------|---------|-------|
-| **🧠 CPU** | All platforms | All platforms | All platforms | ✅ | ✅ |
-| **🎮 GPU** | — | Vulkan: All platforms<br>ROCm: Selected AMD platforms* | — | ✅ | ✅ |
-| **🤖 NPU** | AMD Ryzen™ AI 300 series | — | — | ✅ | — |
+| Hardware | Engine: OGA | Engine: llamacpp | Engine: FLM | Windows | Linux | macOS |
+|----------|-------------|------------------|------------|---------|-------|-------|
+| **🧠 CPU** | All platforms | All platforms | - | ✅ | ✅ | ✅ |
+| **🎮 GPU** | — | Vulkan: All platforms<br>ROCm: Selected AMD platforms*<br>Metal: Apple Silicon | — | ✅ | ✅ | ✅ |
+| **🤖 NPU** | AMD Ryzen™ AI 300 series | — | Ryzen™ AI 300 series | ✅ | — | — |
 
 <details>
 <summary><small><i>* See supported AMD ROCm platforms</i></small></summary>

{lemonade_sdk-8.1.9.dist-info → lemonade_sdk-8.1.11.dist-info}/RECORD

@@ -4,17 +4,17 @@ lemonade/cache.py,sha256=5iZbk273TiTMqK_vdzPOPYTo6VsWW2gNByOISA9zi1w,3002
 lemonade/cli.py,sha256=qU5bW7RQAUKNSpvrhVyzn68NMxyi-336Ke_JU4bsv1Q,5708
 lemonade/sequence.py,sha256=KSH7BPsiyDKsOsg_ziQKEGsDwMmuO_YbgPRBxkZd0pw,13267
 lemonade/state.py,sha256=sdSezla7Cd7KYL90xY3p9kcNV4ndSyN6UvNLOr3vBMA,5261
-lemonade/version.py,sha256=nNxNovCuweKgLzK71oDtPCagHKqrBZW7fNRdrO5VFWA,22
+lemonade/version.py,sha256=et6OH4dSRF6oHhfToZjlMBObl75tflM6mBJXEWJArps,23
 lemonade/common/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 lemonade/common/build.py,sha256=zTb0m1-kuUx6zw5QHp2SNnVuN6jOTMQ2FCdj9iH374U,6140
 lemonade/common/cli_helpers.py,sha256=hjBfXrTtFl8gmCFlL-ksviXR0mOcdPtTWVNKoEp3PG4,4993
 lemonade/common/exceptions.py,sha256=w83sVKmL1QXoJlGjj_bRyjIBMhlMqdVQy_FEOTu2YQI,2050
 lemonade/common/filesystem.py,sha256=QV3cHhKNu-7W2rr8wZ4JQfD2rP_5T2Js7jiDQBYWHVQ,12142
-lemonade/common/inference_engines.py,sha256=pJxn0zOf3gEmjGAIWXNdCibfzarzc7LRbZjoQyygkcU,12591
+lemonade/common/inference_engines.py,sha256=3bUGQe9wtfTiwt8kvI_ry077uyc9lid2G1fJX95kN1A,12969
 lemonade/common/network.py,sha256=qXpUjDYQEYM_gH3JwTtU-pu_yCKcaa1IeohJRPy91-A,2903
 lemonade/common/printing.py,sha256=GFFzrXIineIOMa9yu0lo5sL4j6A5BBg_T9aUCdP-juw,3229
 lemonade/common/status.py,sha256=xSOZN508cdRtrs1HVyr9zmASYg69EsZBLSs0lroLoCM,16519
-lemonade/common/system_info.py,sha256=pn-k3zMQCbt5cu3aHXa4cENgrubOK97gs9PYdGPsFXA,28405
+lemonade/common/system_info.py,sha256=Msa0pCSj3ZN3nerjY8wdqjjJLg6GPhbWf2htSNcFIHc,49607
 lemonade/common/test_helpers.py,sha256=Gwk-pa_6xYAo2oro-2EJNfuouAfw8k_brCbcMC-E-r0,758
 lemonade/profilers/__init__.py,sha256=JKVonvJ4XZ9_6sKXPWsiMLQCNyzQOxhQw5BEHR1qOfU,31
 lemonade/profilers/agt_power.py,sha256=t_37VEg8LPapjSKSjJln-jFznZtTIf5UpzlAXcVGOrc,16771
@@ -31,12 +31,14 @@ lemonade/tools/mmlu.py,sha256=c2QaIMDzjqxCvgHlMXmy_dP1sAFkwkDxL7RO2nogI6s,11071
 lemonade/tools/perplexity.py,sha256=eiaTZ3yhqF2pfwOffVbKKJLwjSri7Im2pC-tBJr7LLU,5638
 lemonade/tools/prompt.py,sha256=PyLksp1k8jsZsU7XBRK61k1DUHhbdLa20h-AP8Noh3w,9011
 lemonade/tools/tool.py,sha256=UsxVYukfm_iM3BfeGYPZxQlTK5UfDfDOl3RIyLr8A1Y,13256
+lemonade/tools/flm/__init__.py,sha256=NQ4CEzJZGS_VvxPMlfrK4Dcx48bQSoUR4iG8e7yZjas,46
+lemonade/tools/flm/utils.py,sha256=hHjSiRlkw239n03CyZeRQomxtmsJptM7m5M3CNnPlqo,8126
 lemonade/tools/huggingface/bench.py,sha256=-mTfldCtquL4mspq8ykVwDc9Mut5Ecv_jHJnSb0CYGE,6734
 lemonade/tools/huggingface/load.py,sha256=KsSGOBBD-tNEIfYC8mCWV_jpnkjHMhN3juVmC1Ln4uQ,7745
 lemonade/tools/huggingface/utils.py,sha256=j1S-IgjDsznUIVwkHSqqChmFyqIx9f3WcEelzohWwvU,13955
 lemonade/tools/llamacpp/bench.py,sha256=1fkE02ecg-jRk92i5dTAXz6re14WH8bd-Z9l-m3lbDA,4844
 lemonade/tools/llamacpp/load.py,sha256=DFCvQN548Ch9H8U_rHOiYviinzw6vixb5-V7xLj7XE4,6499
-lemonade/tools/llamacpp/utils.py,sha256=nl11DRJCzcDV7OcAJtA0YRVxStdBvcnLZbhDhUrN1So,33041
+lemonade/tools/llamacpp/utils.py,sha256=JpI9McEYbrZXQXb0Wo7EoQ8-0LLmmZuwbgGSuuYTiyQ,35221
 lemonade/tools/oga/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 lemonade/tools/oga/bench.py,sha256=PJXv4UchcS2YPwijNzef8DY4DSAKYxIYY1ycHuH3T34,5005
 lemonade/tools/oga/load.py,sha256=x-A-nhoni-WyDpVCLcWRAMfs5ouac9MJzxT-rsnLPw8,34226
@@ -45,33 +47,36 @@ lemonade/tools/report/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3h
 lemonade/tools/report/llm_report.py,sha256=bVHhwCINA-Ok2EdSwAsLubsc83N3KWOVuwTguw7jDcE,6676
 lemonade/tools/report/table.py,sha256=Kv_Epd8a6KIrdzSC2EgIl6uTKw7E5eMq10Tg16O0WxM,27996
 lemonade/tools/server/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-lemonade/tools/server/llamacpp.py,sha256=8HdTkrU2ht8L1ldXqkfYfYhXiA8TvySuaslinAMqr-c,9002
-lemonade/tools/server/serve.py,sha256=gwKPs-r-bfQsyCkVxzVyTp0SyNYVqqfMeeUFQ9TivwM,61638
+lemonade/tools/server/flm.py,sha256=lErpSYLIB6vyVavDd5c-XOz-85m8yPPlyHhiceKWf6c,4119
+lemonade/tools/server/llamacpp.py,sha256=dSab9hR_CcctmU8HeQYQ7U0XCbAQNd_QGZUt6q7cxHA,9952
+lemonade/tools/server/serve.py,sha256=ONdONSN0Noh2c65Z1m9zfQuHIDgb1c42Do0ao2MG534,68771
 lemonade/tools/server/tool_calls.py,sha256=xrAlQwKG-nv2xLlf8f9CDSaUbyMn8ZtHkds9iZLG9K8,5230
-lemonade/tools/server/tray.py,sha256=a9z6hdqlfj91H00j6hAExRPQkzWHhE3dnqSumzEgq0U,19599
-lemonade/tools/server/webapp.py,sha256=8Das5yXOaSBLZmSZ_eddJajQFxBhvl5D6GI_hHlGbE0,1040
-lemonade/tools/server/wrapped_server.py,sha256=DlzsGUwLQzjOFRfTTxfnhvmM_9lvAki96jWIAz7Czds,16713
+lemonade/tools/server/tray.py,sha256=EFnSc2Ra4owiHVz6ykoMhxi2fYqZAK1g21AynAYBiyk,24426
+lemonade/tools/server/webapp.py,sha256=GGSVIzN19C2ZaadOEPBg_D7Lt0PuF339NuWwjMPfZu8,1225
+lemonade/tools/server/wrapped_server.py,sha256=uh7ifrRX1Hx0IuRwZRCGPyQOukitE7kKQipCCz0bSGA,19844
 lemonade/tools/server/static/favicon.ico,sha256=hMmP9qGJNeZ0mFS86JIqPbZstXMZn0Z76_HfHQpREAU,126745
-lemonade/tools/server/static/styles.css,sha256=SYEK4rC-MdpkTj31gxNl9Kb3hCNd1Fpq-EGnRMTjVe8,45362
-lemonade/tools/server/static/webapp.html,sha256=j7A8SOwbY_GfOSkOMV3JvXhOKY1iG70JYYuA3WdoWSQ,17856
-lemonade/tools/server/static/js/chat.js,sha256=XpQSIn1TUra26tu2CtDyOayhXAbUEqzBK0oGtkCAu-s,39162
+lemonade/tools/server/static/logs.html,sha256=BSpdRJ8XJLStpD7XijXLbeeDaVW-FrJbI29zHNBxvYM,1321
+lemonade/tools/server/static/styles.css,sha256=GYJgRtlZSgz3pShDeuatu-J9TpVwGgOjOcIWwNnWYck,49100
+lemonade/tools/server/static/webapp.html,sha256=QQRMMMf8fbtJfGZYqBaRVd0-bENmdPfkVo8sdc4092Q,19151
+lemonade/tools/server/static/js/chat.js,sha256=jxyMyu4MfvI2YmsMbJQ8ZwDNBnLzu2nbjm-qLfgWSNI,42182
 lemonade/tools/server/static/js/model-settings.js,sha256=JXHeG7xVrRU181Hj7CZflERAi1Z6t-qwYFR4aH5nf5I,5820
-lemonade/tools/server/static/js/models.js,sha256=7bCJbvS8FWpj6f1ZOwM8pt6UOQueuiOGLG79qrs-C-A,35872
-lemonade/tools/server/static/js/shared.js,sha256=mD03xqyMH1iQwH4pOq4IpDDaAX0z7YZY71gD8gufRAg,17487
+lemonade/tools/server/static/js/models.js,sha256=es3LwrU49UtoC59e-AusUpdSXQnIRNsZvw8f05keAM0,37856
+lemonade/tools/server/static/js/shared.js,sha256=NVu7lXotLnVVh1x_eXqxX1bLzYyW_eMQttOZ9f0RqUQ,17591
+lemonade/tools/server/utils/macos_tray.py,sha256=xwHW44ZN5hDVlJcwIpHHfqn4VRXWxXHuDACaT-ZqdO8,7095
 lemonade/tools/server/utils/port.py,sha256=J7-g-Aqygb50jNoHLhhRfBZVM-uhGlcB5-oYBAehvgw,2263
-lemonade/tools/server/utils/system_tray.py,sha256=b9lvNv9chJKQxvmH7qzAuUe6H9HsLu7pdHFqGlAJaL0,12654
 lemonade/tools/server/utils/thread.py,sha256=Z-PDzGcpgfN2qxTmtlROWqrUN0B2fXdPrqo_J10fR_w,2772
+lemonade/tools/server/utils/windows_tray.py,sha256=2z5aTmUPlkT-QfkcfwHsyA6dv6nSNBT0gXUErarhac8,13170
 lemonade_install/__init__.py,sha256=26zohKg2jgr_5y7tObduWMYQg8zCTWMZHL8lfi2zZVQ,40
-lemonade_install/install.py,sha256=Dow7kt-K9WI4PH15hBwkKtOxede3dAaOmH4I1y_P5H4,27008
-lemonade_sdk-8.1.9.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-lemonade_sdk-8.1.9.dist-info/licenses/NOTICE.md,sha256=RSca9LE5e6pvdWA_LXAUCcACIHPmINKqkRX-AVRqBGo,3499
-lemonade_server/cli.py,sha256=IsyWGDIeiN0MnIaBryLhArHhPj1HD52l6028dNHQCKc,19248
-lemonade_server/model_manager.py,sha256=V8QRf1nlh3wAFtUHoSF_JeAXeR7sfaZE1uTfppcIfcw,20492
-lemonade_server/pydantic_models.py,sha256=49MyOlb5feLUlKsGcI75tWaflWckrItqcSVkdCY4e3A,3269
-lemonade_server/server_models.json,sha256=0H_G6Jw6Yuz6t0RZnFnq0SbBCsw_cQLe9j24TkyF2eI,12344
-lemonade_server/settings.py,sha256=6nsmPLFJD-UokQDmlx9ZBYMbpnn48So_PuBGWP7Fmfg,1299
-lemonade_sdk-8.1.9.dist-info/METADATA,sha256=rG009a2eVI1apZTBLwIcyQNwdYuVmX33naZwN8Rhbbs,15022
-lemonade_sdk-8.1.9.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-lemonade_sdk-8.1.9.dist-info/entry_points.txt,sha256=7sRvpNhi1E7amnM7RZo57e8yFF9iA5uuRaIeJ1Xre6w,193
-lemonade_sdk-8.1.9.dist-info/top_level.txt,sha256=10ap5GNiPhalO4V50LRoxA1FqRT9g3Xkia6BITu880k,42
-lemonade_sdk-8.1.9.dist-info/RECORD,,
+lemonade_install/install.py,sha256=p3pYqhUnLQ9JJMcbjlSYDYqN-amnU_535O9Oj1yPbyM,27608
+lemonade_sdk-8.1.11.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+lemonade_sdk-8.1.11.dist-info/licenses/NOTICE.md,sha256=RSca9LE5e6pvdWA_LXAUCcACIHPmINKqkRX-AVRqBGo,3499
+lemonade_server/cli.py,sha256=mzlIqLeGkU96KHuib8lmZn0snU7XjQ9lspOy83dVplo,23401
+lemonade_server/model_manager.py,sha256=6W6_nQea6hLD82Il2o_EgQ7oNaLfICXPKNjZiY9Y1Xk,26331
+lemonade_server/pydantic_models.py,sha256=5U3PZ__UqcWQh-dNXVBc-vyJc6-In2vngZXP9VmiScM,3954
+lemonade_server/server_models.json,sha256=_GxymNW7gBkJcTjzncYUvPUJ8kc-I3qqWwHqPztzYcA,14644
+lemonade_server/settings.py,sha256=JOlZmirUXO9rA6BCODVFwyXrrHtYoH_LiKYm49lGm_c,1260
+lemonade_sdk-8.1.11.dist-info/METADATA,sha256=uLq8pbAD_uriJbpPT25i5gk-Ukq_RbWzzEckgtceFDo,15334
+lemonade_sdk-8.1.11.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+lemonade_sdk-8.1.11.dist-info/entry_points.txt,sha256=7sRvpNhi1E7amnM7RZo57e8yFF9iA5uuRaIeJ1Xre6w,193
+lemonade_sdk-8.1.11.dist-info/top_level.txt,sha256=10ap5GNiPhalO4V50LRoxA1FqRT9g3Xkia6BITu880k,42
+lemonade_sdk-8.1.11.dist-info/RECORD,,

lemonade_server/cli.py

@@ -1,6 +1,7 @@
 import argparse
 import sys
 import os
+import platform
 from typing import Tuple, Optional
 import psutil
 from typing import List
@@ -104,12 +105,34 @@ def serve(
         max_wait_time = 30
         wait_interval = 0.5
         waited = 0
-        while waited < max_wait_time:
-            time.sleep(wait_interval)
-            _, running_port = get_server_info()
-            if running_port is not None:
-                break
-            waited += wait_interval
+
+        if platform.system() == "Darwin":
+            # On macOS, use direct HTTP health check instead of process scanning for better
+            # performance
+            import requests
+
+            while waited < max_wait_time:
+                time.sleep(wait_interval)
+                try:
+                    response = requests.get(
+                        f"http://{host}:{port}/api/v1/health", timeout=1
+                    )
+                    if response.status_code == 200:
+                        break
+                except (
+                    requests.exceptions.ConnectionError,
+                    requests.exceptions.Timeout,
+                ):
+                    pass  # Server not ready yet
+                waited += wait_interval
+        else:
+            # On other platforms, use the existing approach
+            while waited < max_wait_time:
+                time.sleep(wait_interval)
+                _, running_port = get_server_info()
+                if running_port is not None:
+                    break
+                waited += wait_interval
 
         return port, server_thread
 
@@ -176,6 +199,7 @@ def pull(
     checkpoint: Optional[str] = None,
     recipe: Optional[str] = None,
     reasoning: bool = False,
+    vision: bool = False,
     mmproj: str = "",
 ):
     """
@@ -202,6 +226,7 @@ def pull(
                 ("checkpoint", checkpoint),
                 ("recipe", recipe),
                 ("reasoning", reasoning),
+                ("vision", vision),
                 ("mmproj", mmproj),
             ]:
                 if value:
@@ -224,6 +249,7 @@ def pull(
             checkpoint=checkpoint,
             recipe=recipe,
             reasoning=reasoning,
+            vision=vision,
             mmproj=mmproj,
             # The pull command will download an upgraded model if available, even
             # if we already have a local copy of the model
@@ -282,6 +308,10 @@ def run(
     import time
     import os
 
+    # Disable tray on macOS for run command due to threading issues
+    if platform.system() == "Darwin":
+        tray = False
+
     # Start the server if not running
     _, running_port = get_server_info()
     server_previously_running = running_port is not None
@@ -367,6 +397,23 @@ def is_lemonade_server(pid):
     """
     Check whether or not a given PID corresponds to a Lemonade server
     """
+    # macOS only: Self-exclusion to prevent blocking server startup
+    if platform.system() == "Darwin":
+        current_pid = os.getpid()
+        if pid == current_pid:
+            return False
+
+        # Exclude children of current process to avoid detecting status commands
+        try:
+            current_process = psutil.Process(current_pid)
+            child_pids = [
+                child.pid for child in current_process.children(recursive=True)
+            ]
+            if pid in child_pids:
+                return False
+        except (psutil.NoSuchProcess, psutil.AccessDenied):
+            pass
+
     try:
         process = psutil.Process(pid)
 
@@ -375,11 +422,29 @@ def is_lemonade_server(pid):
             if process_name in [  # Windows
                 "lemonade-server-dev.exe",
                 "lemonade-server.exe",
+                "lsdev.exe",
             ] or process_name in [  # Linux
                 "lemonade-server-dev",
                 "lemonade-server",
+                "lsdev",
             ]:
                 return True
+            # macOS only: Python scripts appear as "python3.x", check command line
+            elif process_name.startswith("python") and platform.system() == "Darwin":
+                try:
+                    cmdline = process.cmdline()
+                    if len(cmdline) >= 2:
+                        script_path = cmdline[1]
+                        # Check for various lemonade server command patterns (macOS only)
+                        lemonade_patterns = [
+                            "lemonade-server-dev",
+                            "lemonade-server",
+                            "lsdev",  # Short alias for lemonade-server-dev
+                        ]
+                        if any(pattern in script_path for pattern in lemonade_patterns):
+                            return True
+                except (psutil.AccessDenied, psutil.NoSuchProcess):
+                    pass
             elif "llama-server" in process_name:
                 return False
             if not process.parent():
@@ -397,18 +462,43 @@ def get_server_info() -> Tuple[int | None, int | None]:
     2. The port that Lemonade Server is running on
     """
 
-    # Get all network connections and filter for localhost IPv4 listening ports
+    # Try the global approach first (works on Windows/Linux without permissions)
     try:
         connections = psutil.net_connections(kind="tcp4")
-
         for conn in connections:
             if conn.status == "LISTEN" and conn.laddr and conn.pid is not None:
                 if is_lemonade_server(conn.pid):
                     return conn.pid, conn.laddr.port
-
-    except Exception:
+    except (psutil.AccessDenied, PermissionError):
+        # Global approach needs elevated permissions on macOS, fall back to per-process approach
+        pass
+    except Exception:  # pylint: disable=broad-exception-caught
         pass
 
+    # Per-process approach (macOS only - needs this due to permission requirements)
+    if platform.system() == "Darwin":
+        try:
+            for proc in psutil.process_iter(["pid", "name"]):
+                try:
+                    pid = proc.info["pid"]
+                    if is_lemonade_server(pid):
+                        # Found a lemonade server, check its listening ports
+                        connections = proc.net_connections(kind="inet")
+                        for conn in connections:
+                            if conn.status == "LISTEN" and conn.laddr:
+                                return pid, conn.laddr.port
+                        # If no listening connections found, this process is not actually serving
+                        # Continue looking for other processes
+                except (
+                    psutil.NoSuchProcess,
+                    psutil.AccessDenied,
+                    psutil.ZombieProcess,
+                ):
+                    # Some processes may be inaccessible, continue to next
+                    continue
+        except Exception:  # pylint: disable=broad-exception-caught
+            pass
+
     return None, None
 
 
@@ -423,12 +513,13 @@ def list_models():
 
     # Get all supported models and downloaded models
     supported_models = model_manager.supported_models
+    filtered_models = model_manager.filter_models_by_backend(supported_models)
     downloaded_models = model_manager.downloaded_models
 
     # Filter to only show recommended models
     recommended_models = {
         model_name: model_info
-        for model_name, model_info in supported_models.items()
+        for model_name, model_info in filtered_models.items()
         if model_info.get("suggested", False)
     }
 
@@ -505,7 +596,7 @@ def _add_server_arguments(parser):
         "--llamacpp",
         type=str,
         help="LlamaCpp backend to use",
-        choices=["vulkan", "rocm"],
+        choices=["vulkan", "rocm", "metal"],
         default=DEFAULT_LLAMACPP_BACKEND,
     )
     parser.add_argument(
@@ -518,7 +609,7 @@ def _add_server_arguments(parser):
         default=DEFAULT_CTX_SIZE,
     )
 
-    if os.name == "nt":
+    if os.name == "nt" or platform.system() == "Darwin":
         parser.add_argument(
             "--no-tray",
             action="store_true",
@@ -618,7 +709,7 @@ def main():
 
     args = parser.parse_args()
 
-    if os.name != "nt":
+    if os.name != "nt" and platform.system() != "Darwin":
         args.no_tray = True
 
     if args.version: