v0.12.0

See https://github.com/quic/ai-hub-models/releases/v0.12.0 for changelog.

Signed-off-by: QAIHM Team <[email protected]>

README.md

@@ -357,6 +357,7 @@ Qualcomm® AI Hub Models is licensed under BSD-3. See the [LICENSE file](../LICE
 | [DETR-ResNet50](https://aihub.qualcomm.com/models/detr_resnet50) | [qai_hub_models.models.detr_resnet50](qai_hub_models/models/detr_resnet50/README.md) | ✔️ | ✔️ | ✔️
 | [DETR-ResNet50-DC5](https://aihub.qualcomm.com/models/detr_resnet50_dc5) | [qai_hub_models.models.detr_resnet50_dc5](qai_hub_models/models/detr_resnet50_dc5/README.md) | ✔️ | ✔️ | ✔️
 | [MediaPipe-Face-Detection](https://aihub.qualcomm.com/models/mediapipe_face) | [qai_hub_models.models.mediapipe_face](qai_hub_models/models/mediapipe_face/README.md) | ✔️ | ✔️ | ✔️
+| [MediaPipe-Face-Detection-Quantized](https://aihub.qualcomm.com/models/mediapipe_face_quantized) | [qai_hub_models.models.mediapipe_face_quantized](qai_hub_models/models/mediapipe_face_quantized/README.md) | ✔️ | ✔️ | ✔️
 | [MediaPipe-Hand-Detection](https://aihub.qualcomm.com/models/mediapipe_hand) | [qai_hub_models.models.mediapipe_hand](qai_hub_models/models/mediapipe_hand/README.md) | ✔️ | ✔️ | ✔️
 | [YOLOv8-Detection](https://aihub.qualcomm.com/models/yolov8_det) | [qai_hub_models.models.yolov8_det](qai_hub_models/models/yolov8_det/README.md) | ✔️ | ✔️ | ✔️
 | [YOLOv8-Detection-Quantized](https://aihub.qualcomm.com/models/yolov8_det_quantized) | [qai_hub_models.models.yolov8_det_quantized](qai_hub_models/models/yolov8_det_quantized/README.md) | ✔️ | ✔️ | ✔️

qai_hub_models/_version.py

@@ -2,4 +2,4 @@
 # Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
 # SPDX-License-Identifier: BSD-3-Clause
 # ---------------------------------------------------------------------
-__version__ = "0.11.4"
+__version__ = "0.12.0"

qai_hub_models/models/_shared/llama/app.py

@@ -299,7 +299,9 @@ def generate_output_prompt(
         )
         output_token = _get_tokens_from_logits(output)
         past_key_values = get_past_keyval_with_shift(
-            output[1:], num_of_past_key_heads=self.num_past_key_val_heads
+            output[1:],
+            past_key_start=0,
+            num_of_past_key_heads=self.num_past_key_val_heads,
         ).values()
         output_prompt = self.tokenizer.decode(output_token)
         print()

qai_hub_models/models/_shared/llama/model.py

@@ -33,6 +33,7 @@ def get_past_key_names(
     start: int = 0, end: int = 8, num_of_past_key_heads=32, suffix=""
 ):
     past_key_val_name = []
+
     for i in range(start, end):
         cache_names = [
             f"past_key_{i}_h{j}{suffix}" for j in range(num_of_past_key_heads)
@@ -113,7 +114,10 @@ def load_input_cached_data(
 
 
 def get_past_keyval_with_shift(
-    past_key_vals: List[torch.Tensor], num_of_past_key_heads: int = 32
+    past_key_vals: List[torch.Tensor],
+    past_key_start: int,
+    num_of_past_key_heads: int = 32,
+    new_key_suffix: str = "",
 ):
     """
     Clip past key value to feed next iteration
@@ -122,13 +126,18 @@ def get_past_keyval_with_shift(
     total_key_val = num_of_past_key_heads * 2
     for i in range(0, len(past_key_vals), total_key_val):
         l_num = i // total_key_val
+        past_key_num = l_num + past_key_start
         for j, key in enumerate(past_key_vals[i : i + num_of_past_key_heads]):
-            tg_inputs[f"past_key_{l_num}_h{j}"] = key[:, :, :, 1:].detach()
+            tg_inputs[f"past_key_{past_key_num}_h{j}{new_key_suffix}"] = key[
+                :, :, :, 1:
+            ].detach()
 
         for j, val in enumerate(
             past_key_vals[i + num_of_past_key_heads : i + total_key_val]
         ):
-            tg_inputs[f"past_value_{l_num}_h{j}"] = val[:, :, 1:, :].detach()
+            tg_inputs[f"past_value_{past_key_num}_h{j}{new_key_suffix}"] = val[
+                :, :, 1:, :
+            ].detach()
 
     return tg_inputs
 
@@ -244,12 +253,10 @@ def get_output_names(
         # Clipped hidden layers are named same as first part for all parts
         # Eventually, each split should have respective names.
         # layer_start, layer_end = get_hidden_layer_range_from_split(split_part=split_part, model_split_map=model_split_map)
-        layer_range = end - start
-        output_list = [
-            output_name if output_name else f"layers_{layer_range - 1}_add_out_0"
-        ]
+
+        output_list = [output_name if output_name else f"layers_{end - 1}_add_out_0"]
         output_list += get_past_key_names(
-            0, layer_range, num_of_past_key_heads=past_key_val_heads, suffix="_out"
+            start, end, num_of_past_key_heads=past_key_val_heads, suffix="_out"
         )
         return output_list