Add gpu+npu amd hybrid support

.github/actions/server-testing/action.yml

@@ -15,6 +15,10 @@ inputs:
     required: false
     default: ""
     description: "Location of the OGA for RyzenAI NPU install directory on disk"
+  amd_oga_hybrid:
+    required: false
+    default: ""
+    description: "Location of the OGA for RyzenAI Hybrid install directory on disk"
   hf_token:
     required: false
     default: ""

docs/ort_genai_hybrid.md

@@ -0,0 +1,109 @@
+# Introduction
+
+[onnxruntime-genai (aka OGA)](https://github.com/microsoft/onnxruntime-genai/tree/main?tab=readme-ov-file) is a new framework created by Microsoft for running ONNX LLMs.
+
+## Hybrid instructions
+
+### Warnings
+
+ - The OGA wheels need to be installed in a specific order or you will end up with the wrong packages in your environment. If you see pip dependency errors, please delete your conda env and start over with a fresh environment.
+
+### Requirements
+ - [NPU Drivers (version .237)](https://ryzenai.docs.amd.com/en/latest/inst.html#install-npu-drivers)
+ - [Hybrid LLM artifacts package](https://github.com/aigdat/ryzenai-sw-ea/blob/main/ryzen_ai_13_ga/hybrid-llm-artifacts_1.3.0.zip)
+
+### Installation
+
+1. NOTE: ⚠️ DO THESE STEPS IN EXACTLY THIS ORDER ⚠️
+1. Install `lemonade`:
+    1. Create a conda environment: `conda create -n oga-hybrid python=3.10` (Python 3.10 is required)
+    1. Activate: `conda activate oga-hybrid`
+    1. `cd REPO_ROOT`
+    1. `pip install -e .[llm-oga-hybrid]`
+1. Download required OGA packages
+    1. Access the [Hybrid LLM artifacts package](https://account.amd.com/en/member/ryzenai-sw-ea.html#tabs-a5e122f973-item-4757898120-tab) and  download `hybrid-llm-artifacts_1.3.0.zip` and `onnxruntime_vitisai-1.19.0.dev20241217-cp310-cp310-win_amd64.whl`.
+    1. Copy the `onnxruntime_vitisai-1.19.0.dev20241217-cp310-cp310-win_amd64.whl` file to the `hybrid-llm-artifacts_1.3.0\hybrid-llm-artifacts\onnxruntime_genai\wheel` folder.
+    1. Unzip `hybrid-llm-artifacts_1.3.0.zip`
+    1. Create the system environment variable `AMD_OGA_HYBRID` and set it to the path of the `hybrid-llm-artifacts_1.3.0` folder.
+    1. Restart your terminal
+1. Install the wheels:
+    1. `cd hybrid-llm-artifacts_1.3.0\hybrid-llm-artifacts\onnxruntime_genai\wheel`
+    1. `pip install onnxruntime_genai_directml-0.4.0.dev0-cp310-cp310-win_amd64.whl`
+    1. `pip install onnxruntime_vitisai-1.19.0.dev20241217-cp310-cp310-win_amd64.whl`
+1. Install driver
+    1. Download NPU driver from [NPU Drivers (version .237)](https://ryzenai.docs.amd.com/en/latest/inst.html#install-npu-drivers)
+    1. Unzip `NPU_RAI1.3.zip`
+    1. Right click `kipudrv.inf` and select `Install`
+    1. Check under `Device Manager` to ensure that `NPU Compute Accelerator` is using version `32.0.203.237`.
+
+### Runtime
+
+To test basic functionality, point lemonade to any of the models under [quark-quantized-onnx-hybrid-llms-for-ryzen-ai-1.3](https://huggingface.co/collections/amd/quark-awq-g128-int4-asym-fp16-onnx-hybrid-13-674b307d2ffa21dd68fa41d5):
+
+```
+lemonade -i amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid oga-load --device hybrid --dtype int4 llm-prompt -p "hello whats your name?" --max-new-tokens 15
+```
+
+```
+Building "amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid"
+    ✓ Loading OnnxRuntime-GenAI model
+    ✓ Prompting LLM
+
+amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid:
+        <built-in function input> (executed 1x)
+                Build dir:      C:\Users\ramkr\.cache\lemonade\amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid
+                Status:         Successful build!
+                                Dtype:  int4
+                                Device: hybrid
+                                Response:       hello whats your name? i'm a robot, and i'm here to help you with any questions
+
+
+
+Woohoo! Saved to ~\.cache\lemonade\amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid
+```
+
+To test/use the websocket server:
+
+```
+lemonade -i amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid oga-load --device hybrid --dtype int4 serve --max-new-tokens 50
+```
+
+Then open the address (http://localhost:8000) in a browser and chat with it.
+
+```
+Building "amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid"
+    ✓ Loading OnnxRuntime-GenAI model
+      Launching LLM Server
+
+INFO:     Started server process [8704]
+INFO:     Waiting for application startup.
+INFO:     Application startup complete.
+INFO:     Uvicorn running on http://localhost:8000 (Press CTRL+C to quit)
+INFO:     ::1:57038 - "GET / HTTP/1.1" 200 OK
+INFO:     ('::1', 57042) - "WebSocket /ws" [accepted]
+INFO:     connection open
+```
+
+To run a single MMLU test:
+
+```
+lemonade -i amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid oga-load --device hybrid --dtype int4 accuracy-mmlu --tests management
+```
+
+```
+Building "amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid"
+    ✓ Loading OnnxRuntime-GenAI model
+    ✓ Measuring accuracy with MMLU
+
+amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid:
+        <built-in function input> (executed 1x)
+                Build dir:      C:\Users\ramkr\.cache\lemonade\amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid
+                Status:         Successful build!
+                                Dtype:  int4
+                                Device: hybrid
+                                Mmlu Management Accuracy:       49.515 %
+
+
+
+Woohoo! Saved to ~\.cache\lemonade\amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid
+```

docs/ort_genai_igpu.md

@@ -1,6 +1,6 @@
 # OnnxRuntime GenAI (OGA) for iGPU and CPU
 
-onnxruntime-genai (aka OGA) is a new framework created by Microsoft for running ONNX LLMs: https://github.com/microsoft/onnxruntime-genai/tree/main?tab=readme-ov-file
+[onnxruntime-genai (aka OGA)](https://github.com/microsoft/onnxruntime-genai/tree/main?tab=readme-ov-file) is a new framework created by Microsoft for running ONNX LLMs
 
 ## Installation
 

docs/ort_genai_npu.md

@@ -1,6 +1,6 @@
 # Introduction
 
-onnxruntime-genai (aka OGA) is a new framework created by Microsoft for running ONNX LLMs: https://github.com/microsoft/onnxruntime-genai/tree/main?tab=readme-ov-file
+[onnxruntime-genai (aka OGA)](https://github.com/microsoft/onnxruntime-genai/tree/main?tab=readme-ov-file) is a new framework created by Microsoft for running ONNX LLMs
 
 ## NPU instructions
 
@@ -15,10 +15,10 @@ onnxruntime-genai (aka OGA) is a new framework created by Microsoft for running
     1. Create a conda environment: `conda create -n oga-npu python=3.10` (Python 3.10 is required)
     1. Activate: `conda activate oga-npu`
     1. `cd REPO_ROOT`
-    1. `pip install -e .[oga-npu]`
+    1. `pip install -e .[llm-oga-npu]`
 1. Download required OGA packages
-    1. Access the [AMD RyzenAI EA Lounge](https://account.amd.com/en/member/ryzenai-sw-ea.html#tabs-a5e122f973-item-4757898120-tab) and  download `amd_oga_Oct4_2024.zip` from `Ryzen AI 1.3 EA Release`.
-    1. Unzip `amd_oga_Oct4_2024.zip`
+    1. Access the [AMD RyzenAI EA Lounge](https://account.amd.com/en/member/ryzenai-sw-ea.html#tabs-a5e122f973-item-4757898120-tab) and  download `npu-llm-artifacts_1.3.0.zip` from `Ryzen AI 1.3 Model Release`.
+    1. Unzip `npu-llm-artifacts_1.3.0.zip`
 1. Setup your folder structure:
     1. Copy the `amd_oga` folder from the above zip file, if desired
     1. Create the system environment variable `AMD_OGA` and set it to the path to the `amd_oga` folder
@@ -28,79 +28,80 @@ onnxruntime-genai (aka OGA) is a new framework created by Microsoft for running
     1. `pip install onnxruntime_vitisai-1.20.0-cp310-cp310-win_amd64.whl`
     1. `pip install voe-1.2.0-cp310-cp310-win_amd64.whl`
 1. Install driver
-    1. Access the [AMD RyzenAI EA Lounge](https://account.amd.com/en/member/ryzenai-sw-ea.html#tabs-a5e122f973-item-4757898120-tab) and  download `Win24AIDriver.zip` from `Ryzen AI 1.3 Preview Release`.
-    1. Unzip `Win24AIDriver.zip`
+    1. Download NPU driver from [NPU Drivers (version .237)](https://ryzenai.docs.amd.com/en/latest/inst.html#install-npu-drivers)
+    1. Unzip `NPU_RAI1.3.zip`
     1. Right click `kipudrv.inf` and select `Install`
-    1. Check under `Device Manager` to ensure that `NPU Compute Accelerator` is using version `32.0.203.219`.
+    1. Check under `Device Manager` to ensure that `NPU Compute Accelerator` is using version `32.0.203.237`.
 
 ### Runtime
 
-To test basic functionality, point lemonade to any of the models under [quark-quantized-onnx-llms-for-ryzen-ai-1.3-ea](https://huggingface.co/collections/amd/quark-quantized-onnx-llms-for-ryzen-ai-13-ea-66fc8e24927ec45504381902):
+To test basic functionality, point lemonade to any of the models under [quark_awq_g128_int4_asym_bf16_onnx_npu 1.3](https://huggingface.co/collections/amd/quark-awq-g128-int4-asym-bf16-onnx-npu-13-6759f510b8132db53e044aaf)
 
 ```
-lemonade -i amd/Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix oga-load --device npu --dtype int4 llm-prompt -p "hello whats your name?" --max-new-tokens 15
+lemonade -i amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix --device npu --dtype int4 llm-prompt -p "hello whats your name?" --max-new-tokens 15
 ```
 
 ```
-Building "amd_Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix"
-[Vitis AI EP] No. of Operators :   CPU    73 MATMULNBITS    99
-[Vitis AI EP] No. of Subgraphs :MATMULNBITS    33
+Building "amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-fp16-onnx-hybrid"
     ✓ Loading OnnxRuntime-GenAI model
     ✓ Prompting LLM
 
-amd/Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix:
+amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix:
         <built-in function input> (executed 1x)
-                Build dir:      C:\Users\danie/.cache/lemonade\amd_Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix
+                Build dir:      C:\Users\ramkr\.cache\lemonade\amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix
                 Status:         Successful build!
                                 Dtype:  int4
                                 Device: npu
-                                Response:       hello whats your name?
-Hi, I'm a 21 year old male from the
+                                Response:       hello whats your name? i'm a robot, and i'm here to help you with any questions
+
+
+
+Woohoo! Saved to ~\.cache\lemonade\amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix
 ```
 
 To test/use the websocket server:
 
 ```
-lemonade -i amd/Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix oga-load --device npu --dtype int4 serve --max-new-tokens 50
+lemonade -i amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix --device npu --dtype int4 serve --max-new-tokens 50
 ```
 
 Then open the address (http://localhost:8000) in a browser and chat with it.
 
 ```
-Building "amd_Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix"
-[Vitis AI EP] No. of Operators :   CPU    73 MATMULNBITS    99
-[Vitis AI EP] No. of Subgraphs :MATMULNBITS    33
+Building "amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix"
     ✓ Loading OnnxRuntime-GenAI model
+      Launching LLM Server
 
-
-INFO:     Started server process [27752]
+INFO:     Started server process [8704]
 INFO:     Waiting for application startup.
 INFO:     Application startup complete.
 INFO:     Uvicorn running on http://localhost:8000 (Press CTRL+C to quit)
-INFO:     ::1:54973 - "GET / HTTP/1.1" 200 OK
-INFO:     ('::1', 54975) - "WebSocket /ws" [accepted]
+INFO:     ::1:57038 - "GET / HTTP/1.1" 200 OK
+INFO:     ('::1', 57042) - "WebSocket /ws" [accepted]
 INFO:     connection open
-I'm a newbie here. I'm looking for a good place to buy a domain name. I've been looking around and i've found a few good places.
 ```
 
 To run a single MMLU test:
 
 ```
-lemonade -i amd/Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix oga-load --device npu --dtype int4 accuracy-mmlu --tests management
+lemonade -i amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix oga-load --device npu --dtype int4 accuracy-mmlu --tests management
 ```
 
 ```
-Building "amd_Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix"
-[Vitis AI EP] No. of Operators :   CPU    73 MATMULNBITS    99
-[Vitis AI EP] No. of Subgraphs :MATMULNBITS    33
-    ✓ Loading OnnxRuntime-GenAI model   
-    ✓ Measuring accuracy with MMLU   
+Building "amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix"
+    ✓ Loading OnnxRuntime-GenAI model
+    ✓ Measuring accuracy with MMLU
 
-amd/Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix:
+amd/Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix:
         <built-in function input> (executed 1x)
-                Build dir:      C:\Users\danie/.cache/lemonade\amd_Llama-2-7b-hf-awq-g128-int4-asym-fp32-onnx-ryzen-strix    
+                Build dir:      C:\Users\ramkr\.cache\lemonade\amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix
                 Status:         Successful build!
-                                Dtype:  int4 
-                                Device: npu 
-                                Mmlu Management Accuracy:       56.31 %
+                                Dtype:  int4
+                                Device: npu
+                                Mmlu Management Accuracy:       49.515 %
+
+
+
+Woohoo! Saved to ~\.cache\lemonade\amd_Llama-3.2-1B-Instruct-awq-g128-int4-asym-bf16-onnx-ryzen-strix
+
 ```

setup.py

@@ -100,6 +100,12 @@
             "fastapi",
             "uvicorn[standard]",
         ],
+	"llm-oga-hybrid": [
+            "transformers",
+            "torch",
+            "onnx==1.16.1",
+            "numpy==1.26.4",
+        ],
     },
     classifiers=[],
     entry_points={

src/turnkeyml/llm/tools/ort_genai/oga.py

@@ -12,6 +12,7 @@
 import os
 import time
 import json
+import shutil
 from fnmatch import fnmatch
 from queue import Queue
 from huggingface_hub import snapshot_download
@@ -35,7 +36,13 @@
 oga_model_builder_cache_path = "model_builder"
 
 # Mapping from processor to executiion provider, used in pathnames and by model_builder
-execution_providers = {"cpu": "cpu", "npu": "npu", "igpu": "dml", "cuda": "cuda"}
+execution_providers = {
+    "cpu": "cpu",
+    "npu": "npu",
+    "igpu": "dml",
+    "hybrid": "hybrid",
+    "cuda": "cuda",
+}
 
 
 class OrtGenaiTokenizer(TokenizerAdapter):
@@ -248,7 +255,7 @@ def parser(add_help: bool = True) -> argparse.ArgumentParser:
         parser.add_argument(
             "-d",
             "--device",
-            choices=["igpu", "npu", "cpu", "cuda"],
+            choices=["igpu", "npu", "cpu", "hybrid", "cuda"],
             default="igpu",
             help="Which device to load the model on to (default: igpu)",
         )
@@ -312,8 +319,10 @@ def run(
             "cpu": {"int4": "*/*", "fp32": "*/*"},
             "igpu": {"int4": "*/*", "fp16": "*/*"},
             "npu": {"int4": "amd/**-onnx-ryzen-strix"},
+            "hybrid": {"int4": "amd/**-hybrid"},
             "cuda": {"int4": "*/*", "fp16": "*/*"},
         }
+
         hf_supported = (
             device in hf_supported_models
             and dtype in hf_supported_models[device]
@@ -358,7 +367,7 @@ def run(
                 )
 
             # Download the model from HF
-            if device == "npu":
+            if device == "npu" or device == "hybrid":
 
                 # NPU models on HF are ready to go and HF does its own caching
                 full_model_path = snapshot_download(
@@ -367,6 +376,67 @@ def run(
                 )
                 oga_models_subfolder = None
 
+                if device == "hybrid":
+                    # Locate the directory containing hybrid-llm-artifacts_1.3.0 in the system PATH
+                    hybrid_artifacts_path = None
+                    hybrid_artifacts_path = os.environ.get("AMD_OGA_HYBRID")
+
+                    if hybrid_artifacts_path is None:
+                        raise RuntimeError(
+                            "Could not find hybrid-llm-artifacts_1.3.0 in system PATH. "
+                            "Please ensure it is added to your PATH environment variable."
+                        )
+
+                    if hybrid_artifacts_path:
+                        # Construct the path to onnx_custom_ops.dll
+                        custom_ops_path = os.path.join(
+                            hybrid_artifacts_path,
+                            "hybrid-llm-artifacts",
+                            "onnx_utils",
+                            "bin",
+                            "onnx_custom_ops.dll",
+                        )
+
+                        config_path = os.path.join(full_model_path, "genai_config.json")
+
+                        # Check if the config file exists
+                        if os.path.exists(config_path):
+                            with open(config_path, "r", encoding="utf-8") as f:
+                                config = json.load(f)
+
+                            # Modify the custom_ops_library under decoder -> session_options
+                            if (
+                                "model" in config
+                                and "decoder" in config["model"]
+                                and "session_options" in config["model"]["decoder"]
+                            ):
+                                config["model"]["decoder"]["session_options"][
+                                    "custom_ops_library"
+                                ] = custom_ops_path
+
+                            # Write the changes back to the file
+                            with open(config_path, "w", encoding="utf-8") as f:
+                                json.dump(config, f, indent=4)
+
+                        # Copy DirectML.dll from lib to bin folder
+                        src_dll = os.path.join(
+                            hybrid_artifacts_path,
+                            "hybrid-llm-artifacts",
+                            "onnxruntime_genai",
+                            "lib",
+                            "DirectML.dll",
+                        )
+                        dst_dll = os.path.join(
+                            hybrid_artifacts_path,
+                            "hybrid-llm-artifacts",
+                            "onnx_utils",
+                            "bin",
+                            "DirectML.dll",
+                        )
+
+                        # Create destination directory if it doesn't exist
+                        os.makedirs(os.path.dirname(dst_dll), exist_ok=True)
+                        shutil.copy2(src_dll, dst_dll)
             else:
                 # device is 'cpu' or 'igpu'