chore(dependencies): update requirements for CUDA and related libraries on Linux

demo/README.md

@@ -3,10 +3,10 @@
 These clips are small examples pulled from the Riccio exercise-recognition dataset cache used by the
 Modal training pipeline.
 
-| File | Source label | Router label | Duration | FPS | Resolution | Source path |
-| --- | --- | --- | ---: | ---: | --- | --- |
-| `videos/router_demo_push_up_1.mp4` | `push-up` | `push_up` | 5.005s | 29.970 | 1920x1080 | `/data/raw/riccio/Real-Time Exercise Recognition Dataset/final_kaggle_with_additional_video/push-up/push-up_1.mp4` |
-| `videos/router_demo_unknown_barbell_biceps_curl_1.mp4` | `barbell biceps curl` | `unknown` | 3.804s | 29.970 | 1920x1080 | `/data/raw/riccio/Real-Time Exercise Recognition Dataset/final_kaggle_with_additional_video/barbell biceps curl/barbell biceps curl_1.mp4` |
+| File                                                   | Source label          | Router label | Duration |    FPS | Resolution | Source path                                                                                                                                |
+| ------------------------------------------------------ | --------------------- | ------------ | -------: | -----: | ---------- | ------------------------------------------------------------------------------------------------------------------------------------------ |
+| `videos/router_demo_push_up_1.mp4`                     | `push-up`             | `push_up`    |   5.005s | 29.970 | 1920x1080  | `/data/raw/riccio/Real-Time Exercise Recognition Dataset/final_kaggle_with_additional_video/push-up/push-up_1.mp4`                         |
+| `videos/router_demo_unknown_barbell_biceps_curl_1.mp4` | `barbell biceps curl` | `unknown`    |   3.804s | 29.970 | 1920x1080  | `/data/raw/riccio/Real-Time Exercise Recognition Dataset/final_kaggle_with_additional_video/barbell biceps curl/barbell biceps curl_1.mp4` |
 
 The second clip is an unsupported exercise class in the current router contract. It is intentionally
 mapped to `unknown` so the demo includes both a supported exercise and a rejection-class example.

docs/custom-data-collection-guide.md

@@ -8,12 +8,12 @@ each video into the same 30-frame pose-window features used at inference time.
 
 Use these router labels:
 
-| Router label | Folder names accepted by the normalizer | Notes |
-| --- | --- | --- |
-| `squat` | `squat` | Supported exercise class |
-| `push_up` | `push-up`, `push_up`, `pushup` | Supported exercise class |
-| `shoulder_press` | `shoulder press`, `shoulder_press`, `overhead press` | Supported exercise class |
-| `unknown` | `unknown`, unsupported exercise folders, setup/idle/stretching clips | Rejection class |
+| Router label     | Folder names accepted by the normalizer                              | Notes                    |
+| ---------------- | -------------------------------------------------------------------- | ------------------------ |
+| `squat`          | `squat`                                                              | Supported exercise class |
+| `push_up`        | `push-up`, `push_up`, `pushup`                                       | Supported exercise class |
+| `shoulder_press` | `shoulder press`, `shoulder_press`, `overhead press`                 | Supported exercise class |
+| `unknown`        | `unknown`, unsupported exercise folders, setup/idle/stretching clips | Rejection class          |
 
 Unsupported exercises from the Riccio dataset, such as `barbell biceps curl` and `hammer curl`, are
 mapped to `unknown`. Use the same convention for custom negative examples.
@@ -56,16 +56,16 @@ Useful `unknown` clips include:
 
 ## Capture Protocol
 
-| Setting | Recommendation |
-| --- | --- |
-| Clip length | 5-15 seconds for router clips; 10-30 seconds if also used for rep-count evaluation |
-| Frame rate | 24-30 FPS preferred; minimum 15 FPS |
-| Resolution | 720p or better preferred; minimum 480x360 |
-| Camera | Static phone or webcam, no panning |
-| Framing | Full body visible for squats/presses; upper body and hips visible for push-ups |
-| View angle | Side or 30-45 degree angle for squats/push-ups; front or slight-front angle for presses |
-| Lighting | Even lighting; avoid strong backlight and heavy shadows |
-| Reps | 3-8 clean reps per supported clip when possible |
+| Setting     | Recommendation                                                                          |
+| ----------- | --------------------------------------------------------------------------------------- |
+| Clip length | 5-15 seconds for router clips; 10-30 seconds if also used for rep-count evaluation      |
+| Frame rate  | 24-30 FPS preferred; minimum 15 FPS                                                     |
+| Resolution  | 720p or better preferred; minimum 480x360                                               |
+| Camera      | Static phone or webcam, no panning                                                      |
+| Framing     | Full body visible for squats/presses; upper body and hips visible for push-ups          |
+| View angle  | Side or 30-45 degree angle for squats/push-ups; front or slight-front angle for presses |
+| Lighting    | Even lighting; avoid strong backlight and heavy shadows                                 |
+| Reps        | 3-8 clean reps per supported clip when possible                                         |
 
 Keep one athlete and one exercise per clip. Avoid transitions between exercises inside a single clip.
 

docs/exercise-router-training-report.md

@@ -8,64 +8,64 @@ The exercise router was trained and evaluated for `squat`, `push_up`, `shoulder_
 `unknown`. The final active artifact is the BiLSTM temporal model. The scikit-learn baseline is
 retained as a reference and fallback artifact.
 
-| Field | Value |
-| --- | --- |
-| Selected model | `temporal.pt` |
-| Selected artifact | `temporal.pt` |
-| Selection rule | Prefer BiLSTM temporal when available; baseline falls back when temporal is missing |
-| Local active path | `models/exercise_router/active/temporal.pt` |
-| Baseline artifact path | `models/exercise_router/active/router.joblib` |
+| Field                  | Value                                                                     |
+| ---------------------- | ------------------------------------------------------------------------- |
+| Selected model         | `temporal.pt`                                                             |
+| Selected artifact      | `temporal.pt`                                                             |
+| Selection rule         | Prefer BiLSTM temporal when available; baseline falls back when missing   |
+| Local active path      | `models/exercise_router/active/temporal.pt`                               |
+| Baseline artifact path | `models/exercise_router/active/router.joblib`                             |
 
 ## Data
 
-| Metric | Value |
-| --- | ---: |
-| Feature examples | 134 |
-| Window count | 2,224 |
-| Failed feature extractions | 0 |
-| Push-up windows | 287 |
-| Shoulder press windows | 646 |
-| Squat windows | 659 |
-| Unknown windows | 632 |
+| Metric                     | Value |
+| -------------------------- | ----: |
+| Feature examples           |   134 |
+| Window count               | 2,224 |
+| Failed feature extractions |     0 |
+| Push-up windows            |   287 |
+| Shoulder press windows     |   646 |
+| Squat windows              |   659 |
+| Unknown windows            |   632 |
 
 Unsupported Riccio dataset classes such as bicep curl are mapped to `unknown`.
 
 ## Training Setup
 
-| Model | Setup |
-| --- | --- |
+| Model    | Setup                                                                                 |
+| -------- | ------------------------------------------------------------------------------------- |
 | Baseline | scikit-learn `HistGradientBoostingClassifier` over engineered 30-frame window vectors |
-| Temporal | PyTorch BiLSTM over 30-frame feature tensors on Modal A10 |
+| Temporal | PyTorch BiLSTM over 30-frame feature tensors on Modal A10                             |
 
 The BiLSTM hyperparameters follow the Riccio exercise-classification paper:
 
-| Hyperparameter | Value |
-| --- | ---: |
-| Epochs | 73 |
-| Hidden units | 73 |
-| Dropout | 0.2174 |
-| Learning rate | 0.0004 |
-| Batch size | 54 |
+| Hyperparameter      |  Value |
+| ------------------- | -----: |
+| Epochs              |     73 |
+| Hidden units        |     73 |
+| Dropout             | 0.2174 |
+| Learning rate       | 0.0004 |
+| Batch size          |     54 |
 | Final training loss | 0.0003 |
 
 ## Model Complexity
 
 Counts were checked from the local trained artifacts with `uv run --extra train`.
 
-| Model | Count type | Value |
-| --- | --- | ---: |
-| Baseline | Neural-network-style trainable parameters | 0 |
-| Baseline | Input features | 2,574 |
-| Baseline | Trees | 800 |
-| Baseline | Total tree nodes | 20,778 |
-| Baseline | Split nodes | 9,989 |
-| Baseline | Leaves | 10,789 |
-| Baseline | Approximate learned scalar state | 35,915 |
-| BiLSTM temporal | Trainable parameters | 294,924 |
-| BiLSTM temporal | Input features per frame | 429 |
-| BiLSTM temporal | Hidden units | 73 |
-| BiLSTM temporal | Layers | 1 |
-| BiLSTM temporal | Output classes | 4 |
+| Model           | Count type                                |   Value |
+| --------------- | ----------------------------------------- | ------: |
+| Baseline        | Neural-network-style trainable parameters |       0 |
+| Baseline        | Input features                            |   2,574 |
+| Baseline        | Trees                                     |     800 |
+| Baseline        | Total tree nodes                          |  20,778 |
+| Baseline        | Split nodes                               |   9,989 |
+| Baseline        | Leaves                                    |  10,789 |
+| Baseline        | Approximate learned scalar state          |  35,915 |
+| BiLSTM temporal | Trainable parameters                      | 294,924 |
+| BiLSTM temporal | Input features per frame                  |     429 |
+| BiLSTM temporal | Hidden units                              |      73 |
+| BiLSTM temporal | Layers                                    |       1 |
+| BiLSTM temporal | Output classes                            |       4 |
 
 The baseline is a tree-based `HistGradientBoostingClassifier`, so it does not have trainable tensor
 parameters in the same sense as a neural network. The BiLSTM parameter count includes both LSTM
@@ -73,10 +73,10 @@ directions and the linear classification head.
 
 ## Training Metrics
 
-| Model | Validation accuracy | Unknown rejection rate |
-| --- | ---: | ---: |
-| Baseline | 0.9933 | Not reported in baseline training stage |
-| BiLSTM temporal | 0.9820 | 0.9921 |
+| Model           | Validation accuracy |                  Unknown rejection rate |
+| --------------- | ------------------: | --------------------------------------: |
+| Baseline        |              0.9933 | Not reported in baseline training stage |
+| BiLSTM temporal |              0.9820 |                                  0.9921 |
 
 ## Selection Evaluation
 
@@ -84,28 +84,28 @@ The final evaluation scored every available trained artifact on the cached route
 The baseline scored slightly higher on this cache, but the active router is BiLSTM so routing uses
 the temporal pose-window sequence directly.
 
-| Model | Artifact | Accuracy | Unknown rejection rate |
-| --- | --- | ---: | ---: |
-| Baseline | `baseline.joblib` | 0.9987 | 0.9984 |
-| BiLSTM temporal | `temporal.pt` | 0.9964 | 0.9984 |
+| Model           | Artifact          | Accuracy | Unknown rejection rate |
+| --------------- | ----------------- | -------: | ---------------------: |
+| Baseline        | `baseline.joblib` |   0.9987 |                 0.9984 |
+| BiLSTM temporal | `temporal.pt`     |   0.9964 |                 0.9984 |
 
 ### Baseline Precision/Recall
 
-| Label | Precision | Recall |
-| --- | ---: | ---: |
-| `squat` | 0.9985 | 0.9985 |
-| `push_up` | 0.9965 | 1.0000 |
-| `shoulder_press` | 0.9985 | 0.9985 |
-| `unknown` | 1.0000 | 0.9984 |
+| Label            | Precision | Recall |
+| ---------------- | --------: | -----: |
+| `squat`          |    0.9985 | 0.9985 |
+| `push_up`        |    0.9965 | 1.0000 |
+| `shoulder_press` |    0.9985 | 0.9985 |
+| `unknown`        |    1.0000 | 0.9984 |
 
 ### BiLSTM Precision/Recall
 
-| Label | Precision | Recall |
-| --- | ---: | ---: |
-| `squat` | 0.9954 | 0.9954 |
-| `push_up` | 0.9931 | 1.0000 |
-| `shoulder_press` | 0.9969 | 0.9938 |
-| `unknown` | 0.9984 | 0.9984 |
+| Label            | Precision | Recall |
+| ---------------- | --------: | -----: |
+| `squat`          |    0.9954 | 0.9954 |
+| `push_up`        |    0.9931 | 1.0000 |
+| `shoulder_press` |    0.9969 | 0.9938 |
+| `unknown`        |    0.9984 | 0.9984 |
 
 ## Confusion Matrices
 
@@ -114,20 +114,20 @@ Rows are true labels. Columns are predicted labels.
 ### Baseline
 
 | True \\ Predicted | `squat` | `push_up` | `shoulder_press` | `unknown` |
-| --- | ---: | ---: | ---: | ---: |
-| `squat` | 658 | 0 | 1 | 0 |
-| `push_up` | 0 | 287 | 0 | 0 |
-| `shoulder_press` | 0 | 1 | 645 | 0 |
-| `unknown` | 1 | 0 | 0 | 631 |
+| ----------------- | ------: | --------: | ---------------: | --------: |
+| `squat`           |     658 |         0 |                1 |         0 |
+| `push_up`         |       0 |       287 |                0 |         0 |
+| `shoulder_press`  |       0 |         1 |              645 |         0 |
+| `unknown`         |       1 |         0 |                0 |       631 |
 
 ### BiLSTM Temporal
 
 | True \\ Predicted | `squat` | `push_up` | `shoulder_press` | `unknown` |
-| --- | ---: | ---: | ---: | ---: |
-| `squat` | 656 | 0 | 2 | 1 |
-| `push_up` | 0 | 287 | 0 | 0 |
-| `shoulder_press` | 3 | 1 | 642 | 0 |
-| `unknown` | 0 | 1 | 0 | 631 |
+| ----------------- | ------: | --------: | ---------------: | --------: |
+| `squat`           |     656 |         0 |                2 |         1 |
+| `push_up`         |       0 |       287 |                0 |         0 |
+| `shoulder_press`  |       3 |         1 |              642 |         0 |
+| `unknown`         |       0 |         1 |                0 |       631 |
 
 ## Reproduction Commands
 

docs/huggingface-router-model-card.md

@@ -74,10 +74,10 @@ The latest training report is included as `training_report.md`.
 
 Summary:
 
-| Model | Artifact | Accuracy | Unknown rejection rate |
-| --- | --- | ---: | ---: |
-| Baseline | `baseline.joblib` | 0.9987 | 0.9984 |
-| BiLSTM temporal | `temporal.pt` | 0.9964 | 0.9984 |
+| Model           | Artifact          | Accuracy | Unknown rejection rate |
+| --------------- | ----------------- | -------: | ---------------------: |
+| Baseline        | `baseline.joblib` |   0.9987 |                 0.9984 |
+| BiLSTM temporal | `temporal.pt`     |   0.9964 |                 0.9984 |
 
 ## Limitations
 

docs/implementation-plan.md

@@ -121,4 +121,3 @@ Deliverables:
 - Evaluation metrics.
 - Demo videos.
 - README and deployment instructions.
-

docs/pozify-project-documentation.md

@@ -210,7 +210,7 @@ Example:
   "frame_index": 120,
   "timestamp_sec": 4.0,
   "landmarks": {
-    "left_shoulder": {"x": 0.43, "y": 0.32, "z": -0.03, "visibility": 0.99}
+    "left_shoulder": { "x": 0.43, "y": 0.32, "z": -0.03, "visibility": 0.99 }
   },
   "world_landmarks": {},
   "pose_quality": {
@@ -231,7 +231,7 @@ Example:
   "exercise": "push_up",
   "confidence": 0.92,
   "window_predictions": [
-    {"start_sec": 0.0, "end_sec": 1.0, "label": "push_up", "confidence": 0.91}
+    { "start_sec": 0.0, "end_sec": 1.0, "label": "push_up", "confidence": 0.91 }
   ],
   "fallback_required": false
 }

requirements.txt

@@ -48,6 +48,12 @@ colorama==0.4.6 ; sys_platform == 'win32'
     #   tqdm
 contourpy==1.3.3
     # via matplotlib
+cuda-bindings==13.3.1 ; sys_platform == 'linux'
+    # via torch
+cuda-pathfinder==1.5.5 ; sys_platform == 'linux'
+    # via cuda-bindings
+cuda-toolkit==13.0.2 ; sys_platform == 'linux'
+    # via torch
 cycler==0.12.1
     # via matplotlib
 fastapi==0.136.3
@@ -55,7 +61,9 @@ fastapi==0.136.3
     #   gradio
     #   pozify
 filelock==3.29.1
-    # via huggingface-hub
+    # via
+    #   huggingface-hub
+    #   torch
 flatbuffers==25.12.19
     # via mediapipe
 fonttools==4.63.0
@@ -68,6 +76,7 @@ fsspec==2026.4.0
     # via
     #   gradio-client
     #   huggingface-hub
+    #   torch
 gradio==6.17.3
     # via pozify
 gradio-client==2.5.0
@@ -111,7 +120,9 @@ idna==3.18
     #   httpx
     #   yarl
 jinja2==3.1.6
-    # via gradio
+    # via
+    #   gradio
+    #   torch
 joblib==1.5.3
     # via
     #   pozify
@@ -132,6 +143,8 @@ mediapipe==0.10.35
     # via pozify
 modal==1.5.0
     # via pozify
+mpmath==1.3.0
+    # via sympy
 multidict==6.7.1
     # via
     #   aiohttp
@@ -139,6 +152,8 @@ multidict==6.7.1
     #   yarl
 narwhals==2.22.1
     # via scikit-learn
+networkx==3.6.1
+    # via torch
 numpy==1.26.4
     # via
     #   contourpy
@@ -151,6 +166,47 @@ numpy==1.26.4
     #   pozify
     #   scikit-learn
     #   scipy
+nvidia-cublas==13.1.1.3 ; sys_platform == 'linux'
+    # via
+    #   nvidia-cudnn-cu13
+    #   nvidia-cusolver
+    #   torch
+nvidia-cuda-cupti==13.0.85 ; sys_platform == 'linux'
+    # via cuda-toolkit
+nvidia-cuda-nvrtc==13.0.88 ; sys_platform == 'linux'
+    # via
+    #   cuda-toolkit
+    #   nvidia-cublas
+nvidia-cuda-runtime==13.0.96 ; sys_platform == 'linux'
+    # via cuda-toolkit
+nvidia-cudnn-cu13==9.20.0.48 ; sys_platform == 'linux'
+    # via torch
+nvidia-cufft==12.0.0.61 ; sys_platform == 'linux'
+    # via cuda-toolkit
+nvidia-cufile==1.15.1.6 ; sys_platform == 'linux'
+    # via cuda-toolkit
+nvidia-curand==10.4.0.35 ; sys_platform == 'linux'
+    # via cuda-toolkit
+nvidia-cusolver==12.0.4.66 ; sys_platform == 'linux'
+    # via cuda-toolkit
+nvidia-cusparse==12.6.3.3 ; sys_platform == 'linux'
+    # via
+    #   cuda-toolkit
+    #   nvidia-cusolver
+nvidia-cusparselt-cu13==0.8.1 ; sys_platform == 'linux'
+    # via torch
+nvidia-nccl-cu13==2.29.7 ; sys_platform == 'linux'
+    # via torch
+nvidia-nvjitlink==13.0.88 ; sys_platform == 'linux'
+    # via
+    #   cuda-toolkit
+    #   nvidia-cufft
+    #   nvidia-cusolver
+    #   nvidia-cusparse
+nvidia-nvshmem-cu13==3.4.5 ; sys_platform == 'linux'
+    # via torch
+nvidia-nvtx==13.0.85 ; sys_platform == 'linux'
+    # via cuda-toolkit
 opencv-contrib-python==4.11.0.86
     # via mediapipe
 opencv-python-headless==4.11.0.86
@@ -218,6 +274,8 @@ scipy==1.17.1
     # via scikit-learn
 semantic-version==2.10.0
     # via gradio
+setuptools==81.0.0
+    # via torch
 shellingham==1.5.4
     # via typer
 six==1.17.0
@@ -228,6 +286,8 @@ starlette==1.2.1
     # via
     #   fastapi
     #   gradio
+sympy==1.14.0
+    # via torch
 synchronicity==0.12.3
     # via modal
 threadpoolctl==3.6.0
@@ -236,8 +296,12 @@ toml==0.10.2
     # via modal
 tomlkit==0.14.0
     # via gradio
+torch==2.12.0
+    # via pozify
 tqdm==4.68.2
     # via huggingface-hub
+triton==3.7.0 ; sys_platform == 'linux'
+    # via torch
 typer==0.25.1
     # via
     #   gradio
@@ -261,6 +325,7 @@ typing-extensions==4.15.0
     #   pydantic-core
     #   starlette
     #   synchronicity
+    #   torch
     #   typing-inspection
 typing-inspection==0.4.2
     # via