lemonade-sdk 8.0.6__py3-none-any.whl → 8.1.0__py3-none-any.whl

lemonade/tools/server/serve.py

@@ -284,7 +284,7 @@ class Server(ManagementTool):
     def _setup_server_common(
         self,
         port: int,
-        truncate_inputs: bool = False,
+        truncate_inputs: Optional[int] = None,
         log_level: str = DEFAULT_LOG_LEVEL,
         tray: bool = False,
         log_file: str = None,
@@ -295,7 +295,7 @@ class Server(ManagementTool):
 
         Args:
             port: Port number for the server
-            truncate_inputs: Whether to truncate inputs if they exceed max length
+            truncate_inputs: Truncate messages to this length
             log_level: Logging level to configure
             threaded_mode: Whether this is being set up for threaded execution
         """
@@ -372,7 +372,7 @@ class Server(ManagementTool):
         _=None,
         port: int = DEFAULT_PORT,
         log_level: str = DEFAULT_LOG_LEVEL,
-        truncate_inputs: bool = False,
+        truncate_inputs: Optional[int] = None,
         tray: bool = False,
         log_file: str = None,
     ):
@@ -393,7 +393,7 @@ class Server(ManagementTool):
         port: int = DEFAULT_PORT,
         host: str = "localhost",
         log_level: str = "warning",
-        truncate_inputs: bool = False,
+        truncate_inputs: Optional[int] = None,
     ):
         """
         Set up the server for running in a thread.
@@ -1099,29 +1099,20 @@ class Server(ManagementTool):
             )
             self.input_tokens = len(input_ids[0])
 
-        if (
-            self.llm_loaded.max_prompt_length
-            and self.input_tokens > self.llm_loaded.max_prompt_length
-        ):
-            if self.truncate_inputs:
-                # Truncate input ids
-                truncate_amount = self.input_tokens - self.llm_loaded.max_prompt_length
-                input_ids = input_ids[: self.llm_loaded.max_prompt_length]
-
-                # Update token count
-                self.input_tokens = len(input_ids)
-
-                # Show warning message
-                truncation_message = (
-                    f"Input exceeded {self.llm_loaded.max_prompt_length} tokens. "
-                    f"Truncated {truncate_amount} tokens."
-                )
-                logging.warning(truncation_message)
-            else:
-                raise RuntimeError(
-                    f"Prompt tokens ({self.input_tokens}) cannot be greater "
-                    f"than the model's max prompt length ({self.llm_loaded.max_prompt_length})"
-                )
+        if self.truncate_inputs and self.truncate_inputs > self.input_tokens:
+            # Truncate input ids
+            truncate_amount = self.input_tokens - self.truncate_inputs
+            input_ids = input_ids[: self.truncate_inputs]
+
+            # Update token count
+            self.input_tokens = len(input_ids)
+
+            # Show warning message
+            truncation_message = (
+                f"Input exceeded {self.truncate_inputs} tokens. "
+                f"Truncated {truncate_amount} tokens."
+            )
+            logging.warning(truncation_message)
 
         # Log the input tokens early to avoid this not showing due to potential crashes
         logging.debug(f"Input Tokens: {self.input_tokens}")
@@ -1317,7 +1308,7 @@ class Server(ManagementTool):
         self.tokenizer = None
         self.model = None
 
-        default_message = f"model {model_reference} not found"
+        default_message = "see stack trace and error message below"
         if message:
             detail = message
         else:

lemonade/tools/server/static/styles.css

@@ -27,7 +27,6 @@ body {
   min-height: 100vh;
   display: flex;
   flex-direction: column;
-  padding-bottom: 5rem;
 }
 
 body::before {
@@ -102,13 +101,9 @@ body::before {
 }
 
 .site-footer {
-  position: fixed;
-  left: 0;
-  bottom: 0;
-  width: 100%;
   background: transparent;
   padding-top: 0.5rem;
-  z-index: 100;
+  margin-top: auto;
 }
 
 .dad-joke {
@@ -535,6 +530,10 @@ body::before {
   background-color: #ca4747;
 }
 
+.model-label.coding {
+  background-color: #ff6b35;
+}
+
 .model-labels-container {
   display: flex;
   align-items: center;

lemonade/tools/server/static/webapp.html

@@ -109,6 +109,7 @@
                         </label>
                         <select id="register-recipe" name="recipe" required>
                             <option value="llamacpp">llamacpp</option>
+                            <option value="oga-npu">oga-npu</option>
                             <option value="oga-hybrid">oga-hybrid</option>
                             <option value="oga-cpu">oga-cpu</option>
                         </select>
@@ -413,6 +414,8 @@
                     labelClass = 'reasoning';
                 } else if (labelLower === 'reranking') {
                     labelClass = 'reranking';
+                } else if (labelLower === 'coding') {
+                    labelClass = 'coding';
                 }
                 labelSpan.className = `model-label ${labelClass}`;
                 labelSpan.textContent = label;

lemonade/version.py

@@ -1 +1 @@
-__version__ = "8.0.6"
+__version__ = "8.1.0"

lemonade_install/install.py

@@ -24,22 +24,6 @@
 # In any python environment, only one set of artifacts can be installed at a time.
 # Python environments created by Lemonade v6.1.x or earlier will need to be recreated.
 #
-# The Ryzen AI 1.3.0 artifact files use a different directory hierarchy.
-# The Ryzen AI 1.3.0 hybrid artifacts directory hierarchy is:
-#
-#     RYZEN_AI\hybrid\hybrid-llm-artifacts_1.3.0_lounge\hybrid-llm-artifacts\
-#         onnxruntime_genai\lib
-#         onnxruntime_genai\wheel
-#         onnx_utils\bin
-#         eula\eula
-#
-# The Ryzen AI 1.3.0 npu artifacts directory hierarchy is:
-#
-#     RYZEN_AI\npu\amd_oga\
-#         bins\xclbin\stx
-#         libs
-#         wheels
-#
 
 import argparse
 import glob
@@ -56,6 +40,13 @@ import zipfile
 DEFAULT_RYZEN_AI_VERSION = "1.4.0"
 version_info_filename = "version_info.json"
 
+# NPU Driver configuration
+NPU_DRIVER_DOWNLOAD_URL = (
+    "https://account.amd.com/en/forms/downloads/"
+    "ryzenai-eula-public-xef.html?filename=NPU_RAI1.5_280_WHQL.zip"
+)
+REQUIRED_NPU_DRIVER_VERSION = "32.0.203.280"
+
 lemonade_install_dir = Path(__file__).parent.parent.parent
 DEFAULT_QUARK_VERSION = "quark-0.6.0"
 DEFAULT_QUARK_DIR = os.path.join(
@@ -66,14 +57,6 @@ DEFAULT_QUARK_DIR = os.path.join(
 SUPPORTED_RYZEN_AI_SERIES = ["300"]
 
 npu_install_data = {
-    "1.3.0": {
-        "artifacts_zipfile": "ryzen_ai_13_ga/npu-llm-artifacts_1.3.0.zip",
-        "license_file": (
-            "https://account.amd.com/content/dam/account/en/licenses/download/"
-            "amd-end-user-license-agreement.pdf"
-        ),
-        "license_tag": "Beta ",
-    },
     "1.4.0": {
         "artifacts_zipfile": (
             "https://www.xilinx.com/bin/public/openDownload?"
@@ -88,17 +71,6 @@ npu_install_data = {
 }
 
 hybrid_install_data = {
-    "1.3.0": {
-        "artifacts_zipfile": (
-            "https://www.xilinx.com/bin/public/openDownload?"
-            "filename=hybrid-llm-artifacts_1.3.0_012725.zip"
-        ),
-        "license_file": (
-            "https://www.xilinx.com/bin/public/openDownload?"
-            "filename=AMD%20End%20User%20License%20Agreement.pdf"
-        ),
-        "license_tag": "",
-    },
     "1.4.0": {
         "artifacts_zipfile": (
             "https://www.xilinx.com/bin/public/openDownload?"
@@ -154,10 +126,7 @@ def get_oga_npu_dir():
     version_info = get_ryzen_ai_version_info()
     version = version_info["version"]
     ryzen_ai_folder = get_ryzen_ai_path()
-    if "1.3.0" in version:
-        npu_dir = os.path.join(ryzen_ai_folder, "npu", "amd_oga")
-    else:
-        npu_dir = os.path.join(ryzen_ai_folder, "npu")
+    npu_dir = os.path.join(ryzen_ai_folder, "npu")
     if not os.path.isdir(npu_dir):
         raise RuntimeError(
             f"The npu artifacts are missing from the Ryzen AI folder {ryzen_ai_folder}. "
@@ -171,15 +140,7 @@ def get_oga_hybrid_dir():
     version_info = get_ryzen_ai_version_info()
     version = version_info["version"]
     ryzen_ai_folder = get_ryzen_ai_path()
-    if "1.3.0" in version:
-        hybrid_dir = os.path.join(
-            ryzen_ai_folder,
-            "hybrid",
-            "hybrid-llm-artifacts_1.3.0_lounge",
-            "hybrid-llm-artifacts",
-        )
-    else:
-        hybrid_dir = os.path.join(ryzen_ai_folder, "hybrid")
+    hybrid_dir = os.path.join(ryzen_ai_folder, "hybrid")
     if not os.path.isdir(hybrid_dir):
         raise RuntimeError(
             f"The hybrid artifacts are missing from the Ryzen AI folder {ryzen_ai_folder}. "
@@ -189,6 +150,37 @@ def get_oga_hybrid_dir():
     return hybrid_dir, version
 
 
+def _get_ryzenai_version_info(device=None):
+    """
+    Centralized version detection for RyzenAI installations.
+    Uses lazy imports to avoid import errors when OGA is not installed.
+    """
+    try:
+        # Lazy import to avoid errors when OGA is not installed
+        from packaging.version import Version
+        import onnxruntime_genai as og
+
+        if Version(og.__version__) >= Version("0.7.0"):
+            oga_path = os.path.dirname(og.__file__)
+            if og.__version__ == "0.7.0.2":
+                return "1.5.0", oga_path
+            else:
+                return "1.4.0", oga_path
+        else:
+            if device == "npu":
+                oga_path, version = get_oga_npu_dir()
+            else:
+                oga_path, version = get_oga_hybrid_dir()
+            return version, oga_path
+    except ImportError as e:
+        raise ImportError(
+            f"{e}\n Please install lemonade-sdk with "
+            "one of the oga extras, for example:\n"
+            "pip install lemonade-sdk[dev,oga-cpu]\n"
+            "See https://lemonade_server.ai/install_options.html for details"
+        ) from e
+
+
 def download_lfs_file(token, file, output_filename):
     """Downloads a file from LFS"""
     import requests
@@ -426,8 +418,6 @@ class Install:
                 "npu",
                 "hybrid",
                 "unified",
-                "npu-1.3.0",
-                "hybrid-1.3.0",
                 "npu-1.4.0",
                 "hybrid-1.4.0",
                 "unified-1.4.0",
@@ -524,25 +514,14 @@ class Install:
         # Install all whl files in the specified wheels folder
         if wheels_full_path is not None:
             print(f"\nInstalling wheels from {wheels_full_path}\n")
-            if version == "1.3.0":
-                # Install one wheel file at a time (1.3.0 npu build only works this way)
-                for file in os.listdir(wheels_full_path):
-                    if file.endswith(".whl"):
-                        install_cmd = (
-                            f"{sys.executable} -m pip install "
-                            f"{os.path.join(wheels_full_path, file)}"
-                        )
-                        print(f"\nInstalling {file} with command {install_cmd}\n")
-                        subprocess.run(install_cmd, check=True, shell=True)
-            else:
-                # Install all the wheel files together, allowing pip to work out the dependencies
-                wheel_files = glob.glob(os.path.join(wheels_full_path, "*.whl"))
-                install_cmd = [sys.executable, "-m", "pip", "install"] + wheel_files
-                subprocess.run(
-                    install_cmd,
-                    check=True,
-                    shell=True,
-                )
+            # Install all the wheel files together, allowing pip to work out the dependencies
+            wheel_files = glob.glob(os.path.join(wheels_full_path, "*.whl"))
+            install_cmd = [sys.executable, "-m", "pip", "install"] + wheel_files
+            subprocess.run(
+                install_cmd,
+                check=True,
+                shell=True,
+            )
 
         # Delete the zip file
         print(f"\nCleaning up, removing {archive_file_path}\n")
@@ -611,10 +590,7 @@ class Install:
         license_file = npu_install_data[version].get("license_file", None)
         license_tag = npu_install_data[version].get("license_tag", None)
         install_dir = os.path.join(ryzen_ai_folder, "npu")
-        if version == "1.3.0":
-            wheels_full_path = os.path.join(install_dir, "amd_oga/wheels")
-        else:
-            wheels_full_path = os.path.join(install_dir, "wheels")
+        wheels_full_path = os.path.join(install_dir, "wheels")
 
         if license_file:
             Install._get_license_acceptance(version, license_file, license_tag, yes)
@@ -641,17 +617,7 @@ class Install:
         license_file = hybrid_install_data[version].get("license_file", None)
         license_tag = hybrid_install_data[version].get("license_tag", None)
         install_dir = os.path.join(ryzen_ai_folder, "hybrid")
-        if version == "1.3.0":
-            wheels_full_path = os.path.join(
-                ryzen_ai_folder,
-                "hybrid",
-                "hybrid-llm-artifacts_1.3.0_lounge",
-                "hybrid-llm-artifacts",
-                "onnxruntime_genai",
-                "wheel",
-            )
-        else:
-            wheels_full_path = os.path.join(install_dir, "wheels")
+        wheels_full_path = os.path.join(install_dir, "wheels")
 
         if license_file:
             Install._get_license_acceptance(version, license_file, license_tag, yes)
@@ -669,6 +635,21 @@ class Install:
         # Check if the processor is supported before proceeding
         check_ryzen_ai_processor()
 
+        warning_msg = (
+            "\n" + "=" * 80 + "\n"
+            "WARNING: IMPORTANT: NEW RYZEN AI 1.5.0 INSTALLATION PROCESS\n"
+            + "=" * 80
+            + "\n"
+            "Starting with Ryzen AI 1.5.0, installation is now available through PyPI.\n"
+            "For new installations, consider using:\n\n"
+            "pip install lemonade-sdk[oga-ryzenai] --extra-index-url https://pypi.amd.com/simple\n\n"
+            "This legacy installation method (lemonade-install --ryzenai) is still\n"
+            "supported for version 1.4.0, but may be deprecated in future releases.\n"
+            + "=" * 80
+            + "\n"
+        )
+        print(warning_msg)
+
         # Delete any previous Ryzen AI installation in this environment
         ryzen_ai_folder = get_ryzen_ai_path(check_exists=False)
         if os.path.exists(ryzen_ai_folder):

{lemonade_sdk-8.0.6.dist-info → lemonade_sdk-8.1.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lemonade-sdk
-Version: 8.0.6
+Version: 8.1.0
 Summary: Lemonade SDK: Your LLM Aide for Validation and Deployment
 Author-email: lemonade@amd.com
 Requires-Python: >=3.10, <3.13
@@ -22,16 +22,15 @@ Requires-Dist: pytz
 Requires-Dist: zstandard
 Requires-Dist: fastapi
 Requires-Dist: uvicorn[standard]
-Requires-Dist: openai>=1.81.0
+Requires-Dist: openai<1.97.1,>=1.81.0
 Requires-Dist: transformers<=4.53.2
 Requires-Dist: jinja2
 Requires-Dist: tabulate
 Requires-Dist: sentencepiece
 Requires-Dist: huggingface-hub==0.33.0
-Provides-Extra: oga-hybrid
-Requires-Dist: onnx==1.16.1; extra == "oga-hybrid"
-Requires-Dist: numpy==1.26.4; extra == "oga-hybrid"
-Requires-Dist: protobuf>=6.30.1; extra == "oga-hybrid"
+Provides-Extra: oga-ryzenai
+Requires-Dist: onnxruntime-genai-directml-ryzenai==0.7.0.2; extra == "oga-ryzenai"
+Requires-Dist: protobuf>=6.30.1; extra == "oga-ryzenai"
 Provides-Extra: oga-cpu
 Requires-Dist: onnxruntime-genai==0.8.2; extra == "oga-cpu"
 Requires-Dist: onnxruntime>=1.22.0; extra == "oga-cpu"
@@ -43,14 +42,32 @@ Requires-Dist: pandas>=1.5.3; extra == "dev"
 Requires-Dist: matplotlib; extra == "dev"
 Requires-Dist: human-eval-windows==1.0.4; extra == "dev"
 Requires-Dist: lm-eval[api]; extra == "dev"
+Provides-Extra: oga-hybrid
+Requires-Dist: lemonade-sdk[oga-ryzenai]; extra == "oga-hybrid"
+Provides-Extra: oga-unified
+Requires-Dist: lemonade-sdk[oga-ryzenai]; extra == "oga-unified"
 Provides-Extra: oga-hybrid-minimal
-Requires-Dist: lemonade-sdk[oga-hybrid]; extra == "oga-hybrid-minimal"
+Requires-Dist: lemonade-sdk[oga-ryzenai]; extra == "oga-hybrid-minimal"
 Provides-Extra: oga-cpu-minimal
 Requires-Dist: lemonade-sdk[oga-cpu]; extra == "oga-cpu-minimal"
+Provides-Extra: oga-npu-minimal
+Requires-Dist: lemonade-sdk[oga-ryzenai]; extra == "oga-npu-minimal"
 Provides-Extra: llm
 Requires-Dist: lemonade-sdk[dev]; extra == "llm"
 Provides-Extra: llm-oga-cpu
 Requires-Dist: lemonade-sdk[dev,oga-cpu]; extra == "llm-oga-cpu"
+Provides-Extra: llm-oga-npu
+Requires-Dist: onnx==1.16.0; extra == "llm-oga-npu"
+Requires-Dist: onnxruntime==1.18.0; extra == "llm-oga-npu"
+Requires-Dist: numpy==1.26.4; extra == "llm-oga-npu"
+Requires-Dist: protobuf>=6.30.1; extra == "llm-oga-npu"
+Requires-Dist: lemonade-sdk[dev]; extra == "llm-oga-npu"
+Provides-Extra: llm-oga-hybrid
+Requires-Dist: onnx==1.16.1; extra == "llm-oga-hybrid"
+Requires-Dist: numpy==1.26.4; extra == "llm-oga-hybrid"
+Requires-Dist: protobuf>=6.30.1; extra == "llm-oga-hybrid"
+Provides-Extra: llm-oga-unified
+Requires-Dist: lemonade-sdk[dev,llm-oga-hybrid]; extra == "llm-oga-unified"
 Provides-Extra: llm-oga-igpu
 Requires-Dist: onnxruntime-genai-directml==0.6.0; extra == "llm-oga-igpu"
 Requires-Dist: onnxruntime-directml<1.22.0,>=1.19.0; extra == "llm-oga-igpu"
@@ -61,16 +78,6 @@ Requires-Dist: onnxruntime-genai-cuda==0.8.2; extra == "llm-oga-cuda"
 Requires-Dist: onnxruntime-gpu>=1.22.0; extra == "llm-oga-cuda"
 Requires-Dist: transformers<=4.51.3; extra == "llm-oga-cuda"
 Requires-Dist: lemonade-sdk[dev]; extra == "llm-oga-cuda"
-Provides-Extra: llm-oga-npu
-Requires-Dist: onnx==1.16.0; extra == "llm-oga-npu"
-Requires-Dist: onnxruntime==1.18.0; extra == "llm-oga-npu"
-Requires-Dist: numpy==1.26.4; extra == "llm-oga-npu"
-Requires-Dist: protobuf>=6.30.1; extra == "llm-oga-npu"
-Requires-Dist: lemonade-sdk[dev]; extra == "llm-oga-npu"
-Provides-Extra: llm-oga-hybrid
-Requires-Dist: lemonade-sdk[dev,oga-hybrid]; extra == "llm-oga-hybrid"
-Provides-Extra: llm-oga-unified
-Requires-Dist: lemonade-sdk[llm-oga-hybrid]; extra == "llm-oga-unified"
 Dynamic: author-email
 Dynamic: description
 Dynamic: description-content-type
@@ -174,7 +181,7 @@ lemonade-server list
 
 ## Model Library
 
-Lemonade supports both GGUF and ONNX models as detailed in the [Supported Configuration](#supported-configurations) section. A list of all built-in models is available [here](https://lemonade-server.ai/docs/server/models/).
+Lemonade supports both GGUF and ONNX models as detailed in the [Supported Configuration](#supported-configurations) section. A list of all built-in models is available [here](https://lemonade-server.ai/docs/server/server_models/).
 
 You can also import custom GGUF and ONNX models from Hugging Face by using our [Model Manager](http://localhost:8000/#model-management) (requires server to be running).
 <p align="center">
@@ -263,7 +270,7 @@ completion = client.chat.completions.create(
 print(completion.choices[0].message.content)
 ```
 
-For more detailed integration instructions, see the [Integration Guide](./server_integration.md).
+For more detailed integration instructions, see the [Integration Guide](./docs/server/server_integration.md).
 
 ## Beyond an LLM Server
 
@@ -272,6 +279,10 @@ The [Lemonade SDK](./docs/README.md) also include the following components:
 - 🐍 **[Lemonade API](./docs/lemonade_api.md)**: High-level Python API to directly integrate Lemonade LLMs into Python applications.
 - 🖥️ **[Lemonade CLI](./docs/dev_cli/README.md)**: The `lemonade` CLI lets you mix-and-match LLMs (ONNX, GGUF, SafeTensors) with prompting templates, accuracy testing, performance benchmarking, and memory profiling to characterize your models on your hardware.
 
+## FAQ
+
+To read our frequently asked questions, see our [FAQ Guide](./docs/faq.md)
+
 ## Contributing
 
 We are actively seeking collaborators from across the industry. If you would like to contribute to this project, please check out our [contribution guide](./docs/contribute.md).

{lemonade_sdk-8.0.6.dist-info → lemonade_sdk-8.1.0.dist-info}/RECORD

@@ -4,14 +4,14 @@ 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=0UVFH05U_SSITAGCT0SSjcWSJnwORBaDn5ZSlvquMo8,22
+lemonade/version.py,sha256=c04nFsyfS0zYoDvZjLO-uEi12TFB5EWSD6fiWiI7OLQ,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=lcmir_pATr71TfSBJoIZEi3G9xyxNwi2_xpPvPD8_xI,12932
-lemonade/common/network.py,sha256=EkWTxjJ-UM469nEJdzEl-5UNO7dPfVfFzU2SVMMPaD0,1425
+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
 lemonade/common/system_info.py,sha256=dOtX8WLHCz1xmURZWnqhDbyNZv_AulrpX_bbI58eHFQ,27084
@@ -32,13 +32,13 @@ lemonade/tools/tool.py,sha256=UsxVYukfm_iM3BfeGYPZxQlTK5UfDfDOl3RIyLr8A1Y,13256
 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=GBUGOrcDUJdREAIM7GGs4VschqUe-mE_1-MbSUaDjic,4776
+lemonade/tools/llamacpp/bench.py,sha256=1fkE02ecg-jRk92i5dTAXz6re14WH8bd-Z9l-m3lbDA,4844
 lemonade/tools/llamacpp/load.py,sha256=SKacK2n8LpC4DN4yALyEpV2c8_sgOv2G7t6Nlyu7XXg,6273
-lemonade/tools/llamacpp/utils.py,sha256=35eir8PKtxVUDehgb2DJ9hUI0uSjijQgoDK6scmLl1E,22390
+lemonade/tools/llamacpp/utils.py,sha256=vHA5kykkdHSsMGmbEA4RyOHr8wFIh1WenfhCvY8WxZs,22445
 lemonade/tools/oga/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 lemonade/tools/oga/bench.py,sha256=PJXv4UchcS2YPwijNzef8DY4DSAKYxIYY1ycHuH3T34,5005
-lemonade/tools/oga/load.py,sha256=XSznW8lOX_KafSq5J5mIBJzj8YJEBpK0RFGcTE1wnE8,28317
-lemonade/tools/oga/utils.py,sha256=2N1htWM8QKp5g8nHPvk-w9ZYknSc6fWGXcACRkhsXic,16465
+lemonade/tools/oga/load.py,sha256=O82ezF7Jhgz3CJrxDWZYqLHyD_0NS1nsvfMWDaaUI4I,33728
+lemonade/tools/oga/utils.py,sha256=Xd7tmNr69u_bCut0hZqA7saUR3NFZlp4bvWo54mOZb0,16918
 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
@@ -47,26 +47,26 @@ lemonade/tools/report/llm_report.py,sha256=bVHhwCINA-Ok2EdSwAsLubsc83N3KWOVuwTgu
 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=OP0j74QcowEu3zFEcrKIsBbGDOFemBXS5F5DC6oQHaI,18853
-lemonade/tools/server/serve.py,sha256=SXc0qSh-jKS72GlUsuksT7Lov8p3FatgbbycNmHsUfM,57465
+lemonade/tools/server/serve.py,sha256=0-NprfsU-YrX8Qsf1atEi6wPJWemrPjHKEBHV69SwCQ,57046
 lemonade/tools/server/tool_calls.py,sha256=xrAlQwKG-nv2xLlf8f9CDSaUbyMn8ZtHkds9iZLG9K8,5230
 lemonade/tools/server/tray.py,sha256=yoGCM8j_2KzPqo-AlYiauWd8QR56yp6jW6HZ9921Ydg,17525
 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=JmH9LRGB4HGNCxH14owBrUNNBlzx3cVvB3JJ-xJqDqc,26453
-lemonade/tools/server/static/webapp.html,sha256=8khNmsiy4UdjJDkJW3cFeJkmXUR8RQucvCuuka5SNrQ,36230
+lemonade/tools/server/static/styles.css,sha256=8wQ5Cg4rbEh03kC8t7ALE7dB20GiD0Pfu5BAxh9hECU,26429
+lemonade/tools/server/static/webapp.html,sha256=KZm1ZFIhQzLT2Y2wy3hFsQxcOxFzv-blaeLzc1ODhb8,36396
 lemonade/tools/server/utils/port.py,sha256=XnIg2qS73QRrsJn6LgHcrJPmku30Tv6vsYcBVMj82K4,2186
 lemonade/tools/server/utils/system_tray.py,sha256=b9lvNv9chJKQxvmH7qzAuUe6H9HsLu7pdHFqGlAJaL0,12654
 lemonade/tools/server/utils/thread.py,sha256=pK9K_6DNWoQ78NArkAX3Ym2WsxLnCs9sKTk6TitlYnI,2804
 lemonade_install/__init__.py,sha256=26zohKg2jgr_5y7tObduWMYQg8zCTWMZHL8lfi2zZVQ,40
-lemonade_install/install.py,sha256=DJWR36QSjZtvEwRjYPNSjhYgoxLjI_6OPrCMZjL0ChY,28263
-lemonade_sdk-8.0.6.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-lemonade_sdk-8.0.6.dist-info/licenses/NOTICE.md,sha256=B8lEqi4QE41J9ljz4Riv2JgHD1v8GCZE6nNBHO3KIA0,2135
-lemonade_server/cli.py,sha256=2Un5uLK04fIxlfcTiZ0T_EWbbaq2tYymkUHNFeuvB7g,16041
-lemonade_server/model_manager.py,sha256=FfF3z4IpMZqMk_yIo2LHiE76xg7ybROHvi6lcx-0gvE,10754
-lemonade_server/pydantic_models.py,sha256=rp_FFhoTwg6jNmgol-kShwffnRDGbt7jTbIeELvgOIo,2876
-lemonade_server/server_models.json,sha256=wVyjusS5KkOhlQIl1tCnTR4YYbVm7mLU2rHSFk_39hI,7890
-lemonade_sdk-8.0.6.dist-info/METADATA,sha256=g7dOWZPRb0PEyK4UZpVBPnm1LGfZGlkjrdcNfJ_DO_g,15230
-lemonade_sdk-8.0.6.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-lemonade_sdk-8.0.6.dist-info/entry_points.txt,sha256=gJppn0ETtXXR6ceKWEIRdk42kMC7ps59EmU3NCPyPUk,144
-lemonade_sdk-8.0.6.dist-info/top_level.txt,sha256=10ap5GNiPhalO4V50LRoxA1FqRT9g3Xkia6BITu880k,42
-lemonade_sdk-8.0.6.dist-info/RECORD,,
+lemonade_install/install.py,sha256=TBX-VwEHcPo4WX0K_12pKKINnIK3o4SUo3L5XjkqEtw,27669
+lemonade_sdk-8.1.0.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+lemonade_sdk-8.1.0.dist-info/licenses/NOTICE.md,sha256=B8lEqi4QE41J9ljz4Riv2JgHD1v8GCZE6nNBHO3KIA0,2135
+lemonade_server/cli.py,sha256=6QJ5fxNLuVUbuHauA5JHXf0H5dqJ5E4GNTo4YoMOJtg,16049
+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=gitKHj_VHANxjtcXeE5zFpukVO0HyEfKhu3ZaZsj2xo,8867
+lemonade_sdk-8.1.0.dist-info/METADATA,sha256=c3JxCUYw5ujhGSb3FX3mG6UmgG5BLqik8a5j4oe8n7o,15712
+lemonade_sdk-8.1.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+lemonade_sdk-8.1.0.dist-info/entry_points.txt,sha256=gJppn0ETtXXR6ceKWEIRdk42kMC7ps59EmU3NCPyPUk,144
+lemonade_sdk-8.1.0.dist-info/top_level.txt,sha256=10ap5GNiPhalO4V50LRoxA1FqRT9g3Xkia6BITu880k,42
+lemonade_sdk-8.1.0.dist-info/RECORD,,

lemonade_server/cli.py

@@ -57,7 +57,7 @@ def serve(
     log_level = log_level if log_level is not None else DEFAULT_LOG_LEVEL
 
     # Hidden environment variable to enable input truncation (experimental feature)
-    truncate_inputs = "LEMONADE_TRUNCATE_INPUTS" in os.environ
+    truncate_inputs = os.environ.get("LEMONADE_TRUNCATE_INPUTS", None)
 
     # Start the server
     serve_kwargs = {

lemonade_server/model_manager.py

@@ -7,6 +7,7 @@ from importlib.metadata import distributions
 from lemonade_server.pydantic_models import PullConfig
 from lemonade.cache import DEFAULT_CACHE_DIR
 from lemonade.tools.llamacpp.utils import parse_checkpoint, download_gguf
+from lemonade.common.network import custom_snapshot_download
 
 USER_MODELS_FILE = os.path.join(DEFAULT_CACHE_DIR, "user_models.json")
 
@@ -175,7 +176,7 @@ class ModelManager:
             if "gguf" in checkpoint_to_download.lower():
                 download_gguf(gguf_model_config.checkpoint, gguf_model_config.mmproj)
             else:
-                huggingface_hub.snapshot_download(repo_id=checkpoint_to_download)
+                custom_snapshot_download(checkpoint_to_download)
 
             # Register the model in user_models.json, creating that file if needed
             # We do this registration after the download so that we don't register
@@ -233,8 +234,8 @@ class ModelManager:
 
         try:
             # Get the local path using snapshot_download with local_files_only=True
-            snapshot_path = huggingface_hub.snapshot_download(
-                repo_id=base_checkpoint, local_files_only=True
+            snapshot_path = custom_snapshot_download(
+                base_checkpoint, local_files_only=True
             )
 
             # Navigate up to the model directory (parent of snapshots directory)

lemonade_server/pydantic_models.py

@@ -1,4 +1,4 @@
-from typing import Optional, Union, List, Any
+from typing import Optional, Union, List
 
 from pydantic import BaseModel
 
@@ -18,9 +18,6 @@ class LoadConfig(BaseModel):
     model_name: str
     checkpoint: Optional[str] = None
     recipe: Optional[str] = None
-    # Indicates the maximum prompt length allowed for that specific
-    # checkpoint + recipe combination
-    max_prompt_length: Optional[int] = None
     # Indicates whether the model is a reasoning model, like DeepSeek
     reasoning: Optional[bool] = False
     # Indicates which Multimodal Projector (mmproj) file to use