lemonade-sdk 8.1.1__py3-none-any.whl → 8.1.3__py3-none-any.whl

lemonade/tools/server/tray.py

@@ -427,7 +427,7 @@ class LemonadeTray(SystemTray):
         Start the uvicorn server.
         """
         self.server = self.server_factory()
-        self.server.uvicorn_server = self.server.run_in_thread()
+        self.server.uvicorn_server = self.server.run_in_thread(self.server.host)
         self.server.uvicorn_server.run()
 
     def run(self):

lemonade/tools/server/utils/port.py

@@ -5,6 +5,7 @@ import importlib
 import asyncio
 from contextlib import asynccontextmanager
 from fastapi import FastAPI
+from lemonade.version import __version__
 
 _lazy_imports = {
     "TextIteratorStreamer": ("transformers", "TextIteratorStreamer"),
@@ -42,8 +43,8 @@ async def lifespan(app: FastAPI):
         logging.info(
             "\n"
             "\n"
-            "🍋  Lemonade Server Ready!\n"
-            f"🍋    Open http://localhost:{app.port} in your browser for:\n"
+            f"🍋  Lemonade Server v{__version__} Ready!\n"
+            f"🍋    Open http://{app.host_}:{app.port} in your browser for:\n"
             "🍋      💬 chat\n"
             "🍋      💻 model management\n"
             "🍋      📄 docs\n"
@@ -52,8 +53,8 @@ async def lifespan(app: FastAPI):
         logging.info(
             "\n"
             "\n"
-            "[Lemonade]  Lemonade Server Ready!\n"
-            f"[Lemonade]    Open http://localhost:{app.port} in your browser for:\n"
+            f"[Lemonade]  Lemonade Server v{__version__} Ready!\n"
+            f"[Lemonade]    Open http://{app.host_}:{app.port} in your browser for:\n"
             "[Lemonade]      chat\n"
             "[Lemonade]      model management\n"
             "[Lemonade]      docs\n"

lemonade/version.py

@@ -1 +1 @@
-__version__ = "8.1.1"
+__version__ = "8.1.3"

{lemonade_sdk-8.1.1.dist-info → lemonade_sdk-8.1.3.dist-info}/METADATA

@@ -1,18 +1,18 @@
 Metadata-Version: 2.4
 Name: lemonade-sdk
-Version: 8.1.1
+Version: 8.1.3
 Summary: Lemonade SDK: Your LLM Aide for Validation and Deployment
 Author-email: lemonade@amd.com
-Requires-Python: >=3.10, <3.13
+Requires-Python: >=3.10, <3.14
 Description-Content-Type: text/markdown
 License-File: LICENSE
 License-File: NOTICE.md
 Requires-Dist: invoke>=2.0.0
-Requires-Dist: onnx<1.18.0,>=1.11.0
+Requires-Dist: onnx==1.18.0
 Requires-Dist: pyyaml>=5.4
 Requires-Dist: typeguard>=2.3.13
 Requires-Dist: packaging>=20.9
-Requires-Dist: numpy<2.0.0
+Requires-Dist: numpy
 Requires-Dist: fasteners
 Requires-Dist: GitPython>=3.1.40
 Requires-Dist: psutil>=6.1.1
@@ -41,9 +41,10 @@ Requires-Dist: accelerate; extra == "dev"
 Requires-Dist: datasets; extra == "dev"
 Requires-Dist: pandas>=1.5.3; extra == "dev"
 Requires-Dist: matplotlib; extra == "dev"
-Requires-Dist: model-generate==1.5.0; (platform_system == "Windows" and python_version == "3.10") and extra == "dev"
 Requires-Dist: human-eval-windows==1.0.4; extra == "dev"
 Requires-Dist: lm-eval[api]; extra == "dev"
+Provides-Extra: model-generate
+Requires-Dist: model-generate==1.5.0; (platform_system == "Windows" and python_version == "3.10") and extra == "model-generate"
 Provides-Extra: oga-hybrid
 Requires-Dist: lemonade-sdk[oga-ryzenai]; extra == "oga-hybrid"
 Provides-Extra: oga-unified
@@ -105,7 +106,7 @@ Dynamic: summary
     <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="docs/README.md#installation" title="Check out our instructions">
-    <img src="https://img.shields.io/badge/Python-3.10%20%7C%203.12-blue?logo=python&logoColor=white" alt="Made with Python" />
+    <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">
     <img src="https://img.shields.io/badge/PRs-welcome-brightgreen.svg" alt="PRs Welcome" />

{lemonade_sdk-8.1.1.dist-info → lemonade_sdk-8.1.3.dist-info}/RECORD

@@ -4,13 +4,13 @@ lemonade/cache.py,sha256=5iZbk273TiTMqK_vdzPOPYTo6VsWW2gNByOISA9zi1w,3002
 lemonade/cli.py,sha256=9Pcs3PcrWC2F8_pcBaz09xHUICIJTvpemBdPGyXkjIk,4395
 lemonade/sequence.py,sha256=KSH7BPsiyDKsOsg_ziQKEGsDwMmuO_YbgPRBxkZd0pw,13267
 lemonade/state.py,sha256=sdSezla7Cd7KYL90xY3p9kcNV4ndSyN6UvNLOr3vBMA,5261
-lemonade/version.py,sha256=8YlEPKK1Cm5T4dPa2BQPpPwVVTzjPLnmqAeNcTb5nOw,22
+lemonade/version.py,sha256=gnc1sclqzDLnQB9vbqA0LgSMz4H-bYCuu--_P-HWhAc,22
 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=OJQcED9P1ZeQ8d11lDMNeAoaFoUuZlsDcwEZXLbqWRg,12579
+lemonade/common/inference_engines.py,sha256=pJxn0zOf3gEmjGAIWXNdCibfzarzc7LRbZjoQyygkcU,12591
 lemonade/common/network.py,sha256=p1lWJkN0H5hCpb4rKi3Zc47W_BRrrm-7ghdTALJLGqU,1944
 lemonade/common/printing.py,sha256=GFFzrXIineIOMa9yu0lo5sL4j6A5BBg_T9aUCdP-juw,3229
 lemonade/common/status.py,sha256=xSOZN508cdRtrs1HVyr9zmASYg69EsZBLSs0lroLoCM,16519
@@ -24,7 +24,7 @@ lemonade/tools/accuracy.py,sha256=9HCmczDngkBUuUrt49d2CkRo4J0qyWoFYs5cj20bGkg,11
 lemonade/tools/adapter.py,sha256=Ex63Y1SPCOSV4M_QtzEn3YVd39d3yew0lpmEFgp8aH4,3169
 lemonade/tools/bench.py,sha256=aN5LMA_EH6-ZhAH3Gf26JYL7s0eKpUd3j-bReRhzvEY,10016
 lemonade/tools/humaneval.py,sha256=JbxuoOzvR4iyxZv4R6MI7a3gUt5ef_Jj6Ie-9VP2wzY,9531
-lemonade/tools/management_tools.py,sha256=U8GaJnjdXyQ9sw8UxBQMc7glpaLciaVphASaQS4kJsA,10202
+lemonade/tools/management_tools.py,sha256=HQBcr7LYuMqVRYQtvnkNpfutBTA7lblszyoAjjVGu1Y,10201
 lemonade/tools/mmlu.py,sha256=c2QaIMDzjqxCvgHlMXmy_dP1sAFkwkDxL7RO2nogI6s,11071
 lemonade/tools/perplexity.py,sha256=eiaTZ3yhqF2pfwOffVbKKJLwjSri7Im2pC-tBJr7LLU,5638
 lemonade/tools/prompt.py,sha256=PyLksp1k8jsZsU7XBRK61k1DUHhbdLa20h-AP8Noh3w,9011
@@ -34,11 +34,11 @@ lemonade/tools/huggingface/load.py,sha256=KsSGOBBD-tNEIfYC8mCWV_jpnkjHMhN3juVmC1
 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=CTWnzbEYGPSbOizF26yCnyNrHDY19pLusU-YyND992s,29070
+lemonade/tools/llamacpp/utils.py,sha256=Auid9FepxwLIgDahaDNIxwz8kP_ap8Opd3eSF6t637g,32336
 lemonade/tools/oga/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 lemonade/tools/oga/bench.py,sha256=PJXv4UchcS2YPwijNzef8DY4DSAKYxIYY1ycHuH3T34,5005
 lemonade/tools/oga/load.py,sha256=6Pf_QrHpIXDbfpTwFNRj4RmWTxI-RImhYuqRvmTVgmY,33722
-lemonade/tools/oga/utils.py,sha256=Xd7tmNr69u_bCut0hZqA7saUR3NFZlp4bvWo54mOZb0,16918
+lemonade/tools/oga/utils.py,sha256=F8UVLKlfYcLa2SUqlehar8-jaX2Aw4u58DjHNNvLdOA,17675
 lemonade/tools/quark/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 lemonade/tools/quark/quark_load.py,sha256=FJ4LJKTToZbHHWVEOBLadae1a3jCnnY4KvXySHbkJMA,5589
 lemonade/tools/quark/quark_quantize.py,sha256=hwoaXhpBIORvJ16MvewphPkaDEQn3BAgXq5o82Gc-_s,16599
@@ -46,27 +46,31 @@ 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=ssqy1bZqF-wptNzKEOj6_9REtCNZyXO8R5vakAtg3R4,27973
 lemonade/tools/server/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-lemonade/tools/server/llamacpp.py,sha256=KZO4npzefvbaPvlZbpCYsdW0tMSfmmupT8gaK9y65I8,17962
-lemonade/tools/server/serve.py,sha256=PAAGowj2Z5AQIW3G1l52taNyf_0U4kRFR3G735M4DsU,55513
+lemonade/tools/server/llamacpp.py,sha256=jVkaPx1ZbHYiJll3wnDR0fh-e0yfg7UB0BXlLWPx4dE,20998
+lemonade/tools/server/serve.py,sha256=3wnB19YThQLHkjbzy7PCWppQY_j5xKB24GcqM8IybxI,58857
 lemonade/tools/server/tool_calls.py,sha256=xrAlQwKG-nv2xLlf8f9CDSaUbyMn8ZtHkds9iZLG9K8,5230
-lemonade/tools/server/tray.py,sha256=qlQKBkQwG9W2v9GTyycvFc12_jly6vPU1uEkrIFBGTs,17624
+lemonade/tools/server/tray.py,sha256=YJ4-vJlM6tJ0ojY_wVM6COuNscETFkQPt-BaNqYa9YQ,17640
 lemonade/tools/server/webapp.py,sha256=8Das5yXOaSBLZmSZ_eddJajQFxBhvl5D6GI_hHlGbE0,1040
 lemonade/tools/server/static/favicon.ico,sha256=hMmP9qGJNeZ0mFS86JIqPbZstXMZn0Z76_HfHQpREAU,126745
-lemonade/tools/server/static/styles.css,sha256=M_JrH_vML65MWun-C8XCvLOFw35qZURSa77Fk4fVngQ,30029
-lemonade/tools/server/static/webapp.html,sha256=oU6FZHGQCq-SoT6VkWObQvYzzNS0ser5Fmqx2j_5jCI,54380
-lemonade/tools/server/utils/port.py,sha256=XnIg2qS73QRrsJn6LgHcrJPmku30Tv6vsYcBVMj82K4,2186
+lemonade/tools/server/static/styles.css,sha256=X_mqf3XCOo_pZEqkDiVzMGCy8ARseEBq5DdGnAdfVk0,43383
+lemonade/tools/server/static/webapp.html,sha256=FX2MZUsljfgxxuF12KBdgvNkso_z-sHewWc0SEGGcGM,18138
+lemonade/tools/server/static/js/chat.js,sha256=BTvREuEt0NrN8qhAuda5tTAoUN6tbsoukevA-zyTrwQ,27193
+lemonade/tools/server/static/js/model-settings.js,sha256=JXHeG7xVrRU181Hj7CZflERAi1Z6t-qwYFR4aH5nf5I,5820
+lemonade/tools/server/static/js/models.js,sha256=bbX7c8B59ioim86T3x9PFESvF8y3cHPYUO6nhc4SCDs,32500
+lemonade/tools/server/static/js/shared.js,sha256=4iqDNWiKEB7eYS4fdnTy-RwO_ksROrLYLmT2YSomG1M,17065
+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_install/__init__.py,sha256=26zohKg2jgr_5y7tObduWMYQg8zCTWMZHL8lfi2zZVQ,40
 lemonade_install/install.py,sha256=Zl_JtEIhbqZZTvxcqtq895IomEN-JNxp9xOZEtahMHQ,28289
-lemonade_sdk-8.1.1.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-lemonade_sdk-8.1.1.dist-info/licenses/NOTICE.md,sha256=B8lEqi4QE41J9ljz4Riv2JgHD1v8GCZE6nNBHO3KIA0,2135
-lemonade_server/cli.py,sha256=CFfhrRgZNJCd0rDRBF3TeS3dMJgwlKGtvT5_kbsWaXk,17316
-lemonade_server/model_manager.py,sha256=O3fIX52AqU0z10WzPmNEA3lQ_KjOqNq_G-SxjwIgEio,10781
-lemonade_server/pydantic_models.py,sha256=qEvF7x7AuHCHMiByVzGGuLdQTNs233Sw9uQq5cpI6is,2721
-lemonade_server/server_models.json,sha256=iag_dG9S1tkHZUhkJmGAfiUJkgEazdQSv7stC1fVAsQ,9741
-lemonade_sdk-8.1.1.dist-info/METADATA,sha256=XT9cwNUAkhwQ6kad6l7t2nj7m8S0t-9GvaFLOMxLCyE,17065
-lemonade_sdk-8.1.1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-lemonade_sdk-8.1.1.dist-info/entry_points.txt,sha256=gJppn0ETtXXR6ceKWEIRdk42kMC7ps59EmU3NCPyPUk,144
-lemonade_sdk-8.1.1.dist-info/top_level.txt,sha256=10ap5GNiPhalO4V50LRoxA1FqRT9g3Xkia6BITu880k,42
-lemonade_sdk-8.1.1.dist-info/RECORD,,
+lemonade_sdk-8.1.3.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+lemonade_sdk-8.1.3.dist-info/licenses/NOTICE.md,sha256=B8lEqi4QE41J9ljz4Riv2JgHD1v8GCZE6nNBHO3KIA0,2135
+lemonade_server/cli.py,sha256=-haIK4Q9cYFwna5-m6vgxW9qMaGW-_lDFB49zXxDk2A,18755
+lemonade_server/model_manager.py,sha256=cFaHJVOsabwekAPryXAPdo6qrXYBD_yht7XPg2QImqc,10791
+lemonade_server/pydantic_models.py,sha256=oTFnDVCax2Gerz7RBJOJF0FVQjKoUPJZbBo-EgogQyk,3161
+lemonade_server/server_models.json,sha256=DAdG4ebIt5Dy5MM3kmXn1pO0XbNMph1gdpzbacBDVuc,11664
+lemonade_sdk-8.1.3.dist-info/METADATA,sha256=3As4CPILSkJVZMKsyqHZX6o9P8aBsixEJuQTtOas25w,17086
+lemonade_sdk-8.1.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+lemonade_sdk-8.1.3.dist-info/entry_points.txt,sha256=7sRvpNhi1E7amnM7RZo57e8yFF9iA5uuRaIeJ1Xre6w,193
+lemonade_sdk-8.1.3.dist-info/top_level.txt,sha256=10ap5GNiPhalO4V50LRoxA1FqRT9g3Xkia6BITu880k,42
+lemonade_sdk-8.1.3.dist-info/RECORD,,

{lemonade_sdk-8.1.1.dist-info → lemonade_sdk-8.1.3.dist-info}/entry_points.txt

@@ -2,3 +2,4 @@
 lemonade = lemonade:lemonadecli
 lemonade-install = lemonade_install:installcli
 lemonade-server-dev = lemonade_server.cli:main
+lsdev = lemonade_server.cli:developer_entrypoint

lemonade_server/cli.py

@@ -4,6 +4,13 @@ import os
 from typing import Tuple, Optional
 import psutil
 from typing import List
+from lemonade_server.pydantic_models import (
+    DEFAULT_PORT,
+    DEFAULT_HOST,
+    DEFAULT_LOG_LEVEL,
+    DEFAULT_LLAMACPP_BACKEND,
+    DEFAULT_CTX_SIZE,
+)
 
 
 # Error codes for different CLI scenarios
@@ -47,6 +54,7 @@ class ModelLoadError(Exception):
 
 def serve(
     port: int = None,
+    host: str = None,
     log_level: str = None,
     tray: bool = False,
     use_thread: bool = False,
@@ -59,26 +67,20 @@ def serve(
 
     # Otherwise, start the server
     print("Starting Lemonade Server...")
-    from lemonade.tools.server.serve import (
-        Server,
-        DEFAULT_PORT,
-        DEFAULT_LOG_LEVEL,
-        DEFAULT_LLAMACPP_BACKEND,
-        DEFAULT_CTX_SIZE,
-    )
+    from lemonade.tools.server.serve import Server
 
     port = port if port is not None else DEFAULT_PORT
+    host = host if host is not None else DEFAULT_HOST
     log_level = log_level if log_level is not None else DEFAULT_LOG_LEVEL
     llamacpp_backend = (
         llamacpp_backend if llamacpp_backend is not None else DEFAULT_LLAMACPP_BACKEND
     )
-
-    # Use ctx_size if provided, otherwise use default
     ctx_size = ctx_size if ctx_size is not None else DEFAULT_CTX_SIZE
 
     # Start the server
     server = Server(
         port=port,
+        host=host,
         log_level=log_level,
         ctx_size=ctx_size,
         tray=tray,
@@ -259,7 +261,9 @@ def delete(model_names: List[str]):
 def run(
     model_name: str,
     port: int = None,
+    host: str = "localhost",
     log_level: str = None,
+    tray: bool = False,
     llamacpp_backend: str = None,
     ctx_size: int = None,
 ):
@@ -270,13 +274,14 @@ def run(
     import time
 
     # Start the server if not running
-    _, port = get_server_info()
-    server_previously_running = port is not None
+    _, running_port = get_server_info()
+    server_previously_running = running_port is not None
     if not server_previously_running:
         port, server_thread = serve(
             port=port,
+            host=host,
             log_level=log_level,
-            tray=True,
+            tray=tray,
             use_thread=True,
             llamacpp_backend=llamacpp_backend,
             ctx_size=ctx_size,
@@ -291,7 +296,7 @@ def run(
     load(model_name, port)
 
     # Open the webapp with the specified model
-    url = f"http://localhost:{port}/?model={model_name}#llm-chat"
+    url = f"http://{host}:{port}/?model={model_name}#llm-chat"
     print(f"You can now chat with {model_name} at {url}")
     webbrowser.open(url)
 
@@ -440,26 +445,67 @@ def list_models():
     print(tabulate(table_data, headers=headers, tablefmt="simple"))
 
 
+def developer_entrypoint():
+    """
+    Developer entry point that starts the server with debug logging
+    Equivalent to running: lemonade-server-dev serve --log-level debug [additional args]
+
+    This function automatically prepends "serve --log-level debug" to any arguments
+    passed to the lsdev command.
+    """
+    # Save original sys.argv
+    original_argv = sys.argv.copy()
+
+    try:
+        # Take any additional arguments passed to lsdev and append them
+        # after "serve --log-level debug"
+        additional_args = sys.argv[1:] if len(sys.argv) > 1 else []
+
+        # Set sys.argv to simulate "serve --log-level debug" + additional args
+        sys.argv = [sys.argv[0], "serve", "--log-level", "debug"] + additional_args
+        main()
+    finally:
+        # Restore original sys.argv
+        sys.argv = original_argv
+
+
 def _add_server_arguments(parser):
     """Add common server arguments to a parser"""
-    parser.add_argument("--port", type=int, help="Port number to serve on")
+
+    parser.add_argument(
+        "--port",
+        type=int,
+        help="Port number to serve on",
+        default=DEFAULT_PORT,
+    )
+    parser.add_argument(
+        "--host",
+        type=str,
+        help="Address to bind for connections",
+        default=DEFAULT_HOST,
+    )
     parser.add_argument(
         "--log-level",
         type=str,
         help="Log level for the server",
         choices=["critical", "error", "warning", "info", "debug", "trace"],
-        default="info",
+        default=DEFAULT_LOG_LEVEL,
     )
     parser.add_argument(
         "--llamacpp",
         type=str,
-        help=f"LlamaCpp backend to use",
+        help="LlamaCpp backend to use",
         choices=["vulkan", "rocm"],
+        default=DEFAULT_LLAMACPP_BACKEND,
     )
     parser.add_argument(
         "--ctx-size",
         type=int,
-        help="Context size for the model (default: 4096 for llamacpp, truncates prompts for other recipes)",
+        help=(
+            f"Context size for the model (default: {DEFAULT_CTX_SIZE} for llamacpp, "
+            "truncates prompts for other recipes)"
+        ),
+        default=DEFAULT_CTX_SIZE,
     )
 
 
@@ -578,6 +624,7 @@ def main():
             sys.exit(ExitCodes.SERVER_ALREADY_RUNNING)
         serve(
             port=args.port,
+            host=args.host,
             log_level=args.log_level,
             tray=not args.no_tray,
             llamacpp_backend=args.llamacpp,
@@ -603,7 +650,9 @@ def main():
         run(
             args.model,
             port=args.port,
+            host=args.host,
             log_level=args.log_level,
+            tray=not args.no_tray,
             llamacpp_backend=args.llamacpp,
             ctx_size=args.ctx_size,
         )

lemonade_server/model_manager.py

@@ -43,7 +43,7 @@ class ModelManager:
                 if "reasoning" in model_info:
                     model_info["labels"] = (
                         ["reasoning"]
-                        if not model_info["labels"]
+                        if not model_info.get("labels", None)
                         else model_info["labels"] + ["reasoning"]
                     )
                     del model_info["reasoning"]

lemonade_server/pydantic_models.py

@@ -1,10 +1,13 @@
+import os
 from typing import Optional, Union, List
 
 from pydantic import BaseModel
 
-# Set to a high number to allow for interesting experiences in real apps
-# Tests should use the max_new_tokens argument to set a lower value
-DEFAULT_MAX_NEW_TOKENS = 1500
+DEFAULT_PORT = int(os.getenv("LEMONADE_PORT", "8000"))
+DEFAULT_HOST = os.getenv("LEMONADE_HOST", "localhost")
+DEFAULT_LOG_LEVEL = os.getenv("LEMONADE_LOG_LEVEL", "info")
+DEFAULT_LLAMACPP_BACKEND = os.getenv("LEMONADE_LLAMACPP", "vulkan")
+DEFAULT_CTX_SIZE = int(os.getenv("LEMONADE_CTX_SIZE", "4096"))
 
 
 class LoadConfig(BaseModel):
@@ -39,6 +42,9 @@ class CompletionRequest(BaseModel):
     logprobs: int | None = False
     stop: list[str] | str | None = None
     temperature: float | None = None
+    repeat_penalty: float | None = None
+    top_k: int | None = None
+    top_p: float | None = None
     max_tokens: int | None = None
 
 
@@ -56,6 +62,9 @@ class ChatCompletionRequest(BaseModel):
     logprobs: int | None = False
     stop: list[str] | str | None = None
     temperature: float | None = None
+    repeat_penalty: float | None = None
+    top_k: int | None = None
+    top_p: float | None = None
     tools: list[dict] | None = None
     max_tokens: int | None = None
     max_completion_tokens: int | None = None
@@ -95,6 +104,9 @@ class ResponsesRequest(BaseModel):
     model: str
     max_output_tokens: int | None = None
     temperature: float | None = None
+    repeat_penalty: float | None = None
+    top_k: int | None = None
+    top_p: float | None = None
     stream: bool = False
 
 

lemonade_server/server_models.json

@@ -114,6 +114,51 @@
         "recipe": "oga-npu",
         "suggested": true
     },
+    "DeepSeek-R1-Distill-Llama-8B-NPU": {
+        "checkpoint": "amd/DeepSeek-R1-Distill-Llama-8B-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": true
+    },
+    "DeepSeek-R1-Distill-Qwen-7B-NPU": {
+        "checkpoint": "amd/DeepSeek-R1-Distill-Qwen-7B-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": false
+    },
+    "DeepSeek-R1-Distill-Qwen-1.5B-NPU": {
+        "checkpoint": "amd/DeepSeek-R1-Distill-Qwen-1.5B-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": false
+    },
+    "Llama-3.2-3B-Instruct-NPU": {
+        "checkpoint": "amd/Llama-3.2-3B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": false
+    },
+    "Llama-3.2-1B-Instruct-NPU": {
+        "checkpoint": "amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": false
+    },
+    "Mistral-7B-v0.3-Instruct-NPU": {
+        "checkpoint": "amd/Mistral-7B-Instruct-v0.3-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": true
+    },
+    "Phi-3.5-Mini-Instruct-NPU": {
+        "checkpoint": "amd/Phi-3.5-mini-instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": true
+    },
+    "ChatGLM-3-6b-Instruct-NPU": {
+        "checkpoint": "amd/chatglm3-6b-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": false
+    },
+    "AMD-OLMo-1B-Instruct-NPU": {
+        "checkpoint": "amd/AMD-OLMo-1B-SFT-DPO-awq-g128-int4-asym-bf16-onnx-ryzen-strix",
+        "recipe": "oga-npu",
+        "suggested": false
+    },
     "Llama-3.2-1B-Instruct-DirectML": {
         "checkpoint": "amd/Llama-3.2-1B-Instruct-dml-int4-awq-block-128-directml",
         "recipe": "oga-igpu",
@@ -223,8 +268,8 @@
         "checkpoint": "unsloth/cogito-v2-preview-llama-109B-MoE-GGUF:Q4_K_M",
         "mmproj": "mmproj-F16.gguf",
         "recipe": "llamacpp",
-        "suggested": true,
-        "labels": ["vision","hot"]
+        "suggested": false,
+        "labels": ["vision"]
     },
     "nomic-embed-text-v1-GGUF": {
         "checkpoint": "nomic-ai/nomic-embed-text-v1-GGUF:Q4_K_S",
@@ -260,7 +305,7 @@
         "checkpoint": "Qwen/Qwen2.5-Coder-32B-Instruct-GGUF:Q4_K_M",
         "recipe": "llamacpp",
         "suggested": true,  
-        "labels": ["reasoning", "coding"]
+        "labels": ["coding"]
     },
     "gpt-oss-120b-GGUF": {
         "checkpoint": "unsloth/gpt-oss-120b-GGUF:Q4_K_M",
@@ -273,5 +318,11 @@
         "recipe": "llamacpp",
         "suggested": true,
         "labels": ["hot", "reasoning"]
+    },
+    "GLM-4.5-Air-UD-Q4K-XL-GGUF": {
+        "checkpoint": "unsloth/GLM-4.5-Air-GGUF:UD-Q4_K_XL",
+        "recipe": "llamacpp",
+        "suggested": true,
+        "labels": ["reasoning","hot"]
     }
 }