Add NPU OGA support for RyzenAI 1.3 EA (#232)

* Add NPU OGA support for RyzenAI 1.3 EA

* retrigger checks

src/turnkeyml/llm/docs/ort_genai_npu.md

@@ -0,0 +1,111 @@
+# 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
+
+## NPU instructions
+
+### Warnings
+
+ - Users have experienced inconsistent results across models and machines. If one model isn't working well on your laptop, try one of the other models.
+ - 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.
+
+### Installation
+
+1. NOTE: ⚠️ DO THESE STEPS IN EXACTLY THIS ORDER ⚠️
+1. Install `lemonade`:
+    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. 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 Preview Release`.
+    1. Unzip `amd_oga_Oct4_2024.zip`
+1. Setup your folder structure:
+    1. Copy all of the content inside `amd_oga` to lemonade's `REPO_ROOT\src\lemonade\tools\ort_genai\models\`
+    1. Move all dlls from `REPO_ROOT\src\lemonade\tools\ort_genai\models\libs` to `REPO_ROOT\src\lemonade\tools\ort_genai\models\`
+1. Install the wheels:
+    1. `cd amd_oga\wheels`
+    1. `pip install onnxruntime_genai-0.5.0.dev0-cp310-cp310-win_amd64.whl`
+    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. Ensure you have access to the models on Hungging Face:
+    1. Ensure you can access the models under [quark-quantized-onnx-llms-for-ryzen-ai-13-ea](https://huggingface.co/collections/amd/quark-quantized-onnx-llms-for-ryzen-ai-13-ea-66fc8e24927ec45504381902) on Hugging Face. Models are gated and you may have to request access.
+    1. Create a Hugging Face Access Token [here](https://huggingface.co/settings/tokens). Ensure you select `Read access to contents of all public gated repos you can access` if creating a finegrained token.
+    1. Set your Hugging Face token as an environment variable: `set HF_TOKEN=<your token>`
+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. 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`.
+
+### Runtime
+
+To test basic functionality, point lemonade to any of the models under under [quark-quantized-onnx-llms-for-ryzen-ai-13-ea](https://huggingface.co/collections/amd/quark-quantized-onnx-llms-for-ryzen-ai-13-ea-66fc8e24927ec45504381902):
+
+```
+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
+```
+
+```
+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
+    ✓ Prompting LLM
+
+amd/Llama-2-7b-hf-awq-g128-int4-asym-fp32-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
+                Status:         Successful build!
+                                Dtype:  int4
+                                Device: npu
+                                Response:       hello whats your name?
+Hi, I'm a 21 year old male from the
+```
+
+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
+```
+
+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
+    ✓ Loading OnnxRuntime-GenAI model
+
+
+INFO:     Started server process [27752]
+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:     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
+```
+
+```
+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   
+
+amd/Llama-2-7b-hf-awq-g128-int4-asym-fp32-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    
+                Status:         Successful build!
+                                Dtype:  int4 
+                                Device: npu 
+                                Mmlu Management Accuracy:       56.31 %
+```

src/turnkeyml/llm/leap.py

@@ -90,7 +90,7 @@ def from_pretrained(
 
         return state.model, state.tokenizer
 
-    if recipe == "hf-dgpu":
+    elif recipe == "hf-dgpu":
         # Huggingface Transformers recipe for discrete GPU (Nvidia, Instinct, Radeon)
 
         import torch

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

@@ -6,7 +6,9 @@
 import os
 import time
 import json
+from fnmatch import fnmatch
 from queue import Queue
+from huggingface_hub import snapshot_download, login
 import onnxruntime_genai as og
 from turnkeyml.state import State
 from turnkeyml.tools import FirstTool
@@ -18,7 +20,6 @@
 )
 from turnkeyml.llm.cache import Keys
 
-
 class OrtGenaiTokenizer(TokenizerAdapter):
     def __init__(self, model: og.Model):
         # Initialize the tokenizer and produce the initial tokens.
@@ -74,9 +75,9 @@ def __init__(self, input_folder):
         self.config = self.load_config(input_folder)
 
     def load_config(self, input_folder):
-        config_path = os.path.join(input_folder, 'genai_config.json')
+        config_path = os.path.join(input_folder, "genai_config.json")
         if os.path.exists(config_path):
-            with open(config_path, 'r', encoding='utf-8') as f:
+            with open(config_path, "r", encoding="utf-8") as f:
                 return json.load(f)
         return None
 
@@ -99,21 +100,23 @@ def generate(
         max_length = len(input_ids) + max_new_tokens
 
         params.input_ids = input_ids
-        if self.config and 'search' in self.config:
-            search_config = self.config['search']
+        if self.config and "search" in self.config:
+            search_config = self.config["search"]
             params.set_search_options(
-                do_sample=search_config.get('do_sample', do_sample),
-                top_k=search_config.get('top_k', top_k),
-                top_p=search_config.get('top_p', top_p),
-                temperature=search_config.get('temperature', temperature),
+                do_sample=search_config.get("do_sample", do_sample),
+                top_k=search_config.get("top_k", top_k),
+                top_p=search_config.get("top_p", top_p),
+                temperature=search_config.get("temperature", temperature),
                 max_length=max_length,
                 min_length=0,
-                early_stopping=search_config.get('early_stopping', False),
-                length_penalty=search_config.get('length_penalty', 1.0),
-                num_beams=search_config.get('num_beams', 1),
-                num_return_sequences=search_config.get('num_return_sequences', 1),
-                repetition_penalty=search_config.get('repetition_penalty', 1.0),
-                past_present_share_buffer=search_config.get('past_present_share_buffer', True),
+                early_stopping=search_config.get("early_stopping", False),
+                length_penalty=search_config.get("length_penalty", 1.0),
+                num_beams=search_config.get("num_beams", 1),
+                num_return_sequences=search_config.get("num_return_sequences", 1),
+                repetition_penalty=search_config.get("repetition_penalty", 1.0),
+                past_present_share_buffer=search_config.get(
+                    "past_present_share_buffer", True
+                ),
                 # Not currently supported by OGA
                 # diversity_penalty=search_config.get('diversity_penalty', 0.0),
                 # no_repeat_ngram_size=search_config.get('no_repeat_ngram_size', 0),
@@ -192,6 +195,7 @@ class OgaLoad(FirstTool):
         llama_2 = "meta-llama/Llama-2-7b-chat-hf"
         phi_3_mini_4k = "microsoft/Phi-3-mini-4k-instruct"
         phi_3_mini_128k = "microsoft/Phi-3-mini-128k-instruct"
+        And models on Hugging Face that follow the "amd/**-onnx-ryzen-strix" pattern
 
     Output:
         state.model: handle to a Huggingface-style LLM loaded on DirectML device
@@ -244,7 +248,7 @@ def run(
         checkpoint = input
 
         # Map of models[device][dtype][checkpoint] to the name of the model folder on disk
-        supported_models = {
+        local_supported_models = {
             "igpu": {
                 "int4": {
                     phi_3_mini_128k: os.path.join(
@@ -261,6 +265,7 @@ def run(
             },
             "npu": {
                 "int4": {
+                    # Legacy RyzenAI 1.2 models for NPU
                     llama_2: "llama2-7b-int4",
                     llama_3: "llama3-8b-int4",
                     qwen_1dot5: "qwen1.5-7b-int4",
@@ -277,28 +282,60 @@ def run(
             },
         }
 
+        hf_supported_models = {"npu": {"int4": "amd/**-onnx-ryzen-strix"}}
+
+        supported_locally = True
         try:
-            dir_name = supported_models[device][dtype][checkpoint]
+            dir_name = local_supported_models[device][dtype][checkpoint]
         except KeyError as e:
-            raise ValueError(
-                "The device;dtype;checkpoint combination is not supported: "
-                f"{device};{dtype};{checkpoint}. The supported combinations "
-                f"are: {supported_models}"
-            ) from e
-
-        model_dir = os.path.join(
-            os.path.dirname(os.path.realpath(__file__)),
-            "models",
-            dir_name,
-        )
+            supported_locally = False
+            hf_supported = (
+                device in hf_supported_models
+                and dtype in hf_supported_models[device]
+                and fnmatch(checkpoint, hf_supported_models[device][dtype])
+            )
+            if not hf_supported:
+                raise ValueError(
+                    "The device;dtype;checkpoint combination is not supported: "
+                    f"{device};{dtype};{checkpoint}. The supported combinations "
+                    f"are: {local_supported_models} for local models and {hf_supported_models}"
+                    " for models on Hugging Face."
+                ) from e
+
+        # Create models dir if it doesn't exist
+        models_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)), "models")
+        if not os.path.exists(models_dir):
+            os.makedirs(models_dir)
+
+        # If the model is supported though Hugging Face, download it
+        if not supported_locally:
+            hf_model_name = checkpoint.split("amd/")[1]
+            dir_name = "_".join(hf_model_name.split("-")[:6]).lower()
+            api_key = os.getenv("HF_TOKEN")
+            login(api_key)
+            snapshot_download(
+                repo_id=checkpoint,
+                local_dir=os.path.join(models_dir, dir_name),
+                ignore_patterns=["*.md", "*.txt"],
+            )
 
-        # The NPU requires the CWD to be in the model folder
         current_cwd = os.getcwd()
         if device == "npu":
-            os.chdir(model_dir)
-            # Required environment variable for NPU
-            os.environ["DOD_ROOT"] = ".\\bins"
+            # Change to the models directory
+            os.chdir(models_dir)
+
+            # Common environment variables for all NPU models
+            os.environ["DD_ROOT"] = ".\\bins"
+            os.environ["DEVICE"] = "stx"
+            os.environ["XLNX_ENABLE_CACHE"] = "0"
+
+            # Phi models require USE_AIE_RoPE=0
+            if "phi-" in checkpoint.lower():
+                os.environ["USE_AIE_RoPE"] = "0"
+            else:
+                os.environ["USE_AIE_RoPE"] = "1"
 
+        model_dir = os.path.join(models_dir, dir_name)
         state.model = OrtGenaiModel(model_dir)
         state.tokenizer = OrtGenaiTokenizer(state.model.model)
         state.dtype = dtype