Phi4 mini instruct QNN recipie (#517)

Co-authored-by: Ronak Mahawar <rmahawar@qti.qualcomm.com>

Author: rM-planet

microsoft-Phi-4-mini-instruct/QAIRT/README.md

@@ -0,0 +1,78 @@
+# Phi-4-mini-instruct Model Optimization
+
+This directory demonstrates the optimization of the [Phi-4-mini-instruct](https://huggingface.co/microsoft/Phi-4-mini-instruct) model using various AIMET quantization techniques.
+
+## Overview
+
+After quantization, the QAIRT GenAIBuilder API is utilized to apply additional model transformations, perform conversion, and compile the model for execution on the HTP backend.
+
+Finally, a prepared QAIRT DLC is encapsulated in an ONNX protobuf and exported to a directory compatible with onnxruntime-genai.
+
+## Requirements
+
+**Validated host configuration:**
+* Ubuntu 22.04
+* Python 3.10.12
+* qairt-dev 0.8.1
+* QAIRT 2.45.40
+
+**Validated target configuration:**
+* HTP backend on SC8480XP
+
+Other configurations may work but have not been validated.
+
+## Preparation Instructions
+
+1. Authenticate with Hugging Face
+
+```bash
+huggingface-cli login  # Recommended: stores credentials securely, avoids shell history
+# Alternative: export HF_TOKEN=<your_hugging_face_token>
+```
+
+2. Prepare Environment
+
+```bash
+pip install --no-deps -r requirements.txt
+pip install --no-build-isolation git+https://github.com/microsoft/Olive.git@f7efd41ab24a2eb07be7edc6d84d0f6304b46598
+pip install --no-deps qairt-dev==0.8.1  # Install the proper qairt-dev version, if not installed
+```
+
+3. Use qairt-vm to install a non-default version of QAIRT and set QAIRT_SDK_ROOT
+
+```bash
+# List available QAIRT SDK versions
+qairt-vm fetch --list
+
+# Download non-default version of QAIRT SDK
+qairt-vm fetch -v <version>
+
+# Set QAIRT_SDK_ROOT to download location of QAIRT SDK
+# By default, /opt/qcom/aistack/qairt/<version>
+# Note: No further QAIRT SDK installation steps are required when using qairt-dev
+export QAIRT_SDK_ROOT=/path/to/qairt/sdk
+```
+
+4. Run Olive recipe
+
+```bash
+# For X Elite:
+olive run --config htp_sc8380xp.json
+```
+
+## Execution Instructions
+
+The output of the above olive recipe is a directory compatible with the following versions of onnxruntime-genai and onnxruntime-qnn.
+
+```bash
+pip install onnxruntime-genai>=0.13
+pip install onnxruntime-qnn>=2.1.0
+```
+
+Please see the following script in the onnxruntime-genai repository for [an example of how to run this model directory](https://github.com/microsoft/onnxruntime-genai/blob/main/examples/python/model-qa.py).
+
+## Known Issues
+
+### `AttributeError: module 'pydantic._internal._typing_extra' has no attribute 'add_module_globals'`
+
+This error can occasionally occur on the first invocation of the recipe. If encountered, re-running the recipe is sufficient as a workaround.

microsoft-Phi-4-mini-instruct/QAIRT/config/mixed_precision_config/exceptions.json

@@ -0,0 +1,70 @@
+{
+    "module_list": [
+        {
+            "module_name": "QuantizedRmsNorm",
+            "exceptions": {
+                "param_exceptions": { "asymmetric": true, "bitwidth": 16 },
+                "input_exceptions": null,
+                "output_exceptions": null
+            }
+        }
+    ],
+    "name_list": [
+        {
+            "module_name": "\\w*model_embed_tokens_Gather",
+            "exceptions": {
+                "param_exceptions": { "bitwidth": 16, "asymmetric": true },
+                "input_exceptions": null,
+                "output_exceptions": null
+            }
+        },
+        {
+            "module_name": "\\w*lm_head_(MatMul|conv_Conv|conv2d_Conv|Conv)",
+            "exceptions": {
+                "param_exceptions": { "bitwidth": 16 },
+                "input_exceptions": null,
+                "output_exceptions": null
+            }
+        },
+        {
+            "module_name": "\\w*norm_(Mul_1|Mul_1.module)",
+            "exceptions": {
+                "param_exceptions": null,
+                "input_exceptions": [ { "input_index": 0, "bitwidth": 16, "asymmetric": true } ],
+                "output_exceptions": null
+            }
+        },
+        {
+            "module_name": "\\w*norm_(Pow|Pow.module|ReduceMean|Add|Sqrt|Div|Mul)",
+            "exceptions": {
+                "param_exceptions": null,
+                "input_exceptions": null,
+                "output_exceptions": [ { "output_index": 0, "enabled": false } ]
+            }
+        },
+        {
+            "module_name": "\\w*self_attn_Concat_1",
+            "exceptions": {
+                "param_exceptions": null,
+                "input_exceptions": null,
+                "output_exceptions": [ { "output_index": 0, "bitwidth": 16, "asymmetric": false } ]
+            }
+        },
+        {
+            "module_name": "\\w*self_attn_Concat_4",
+            "exceptions": {
+                "param_exceptions": null,
+                "input_exceptions": null,
+                "output_exceptions": [ { "output_index": 0, "bitwidth": 16, "asymmetric": false } ]
+            }
+        },
+        {
+            "module_name": "\\w*v_proj_(MatMul|conv_Conv|conv2d_Conv|Conv)(\\.base_layer)?",
+            "exceptions": {
+                "param_exceptions": null,
+                "input_exceptions": null,
+                "output_exceptions": [ { "output_index": 0, "bitwidth": 16, "asymmetric": false } ]
+            }
+        }
+    ]
+}