poly-hammer-worker 0.1.1.dev12__tar.gz → 0.3.8__tar.gz

poly_hammer_worker-0.3.8/.dockerignore

@@ -0,0 +1,35 @@
+# Dependencies
+node_modules
+**/node_modules
+frontend/node_modules
+
+# Build outputs
+dist
+**/dist
+frontend/dist
+
+# Python
+__pycache__
+*.pyc
+.venv
+backend/.venv
+
+# Git
+.git
+.gitignore
+
+# IDE
+.vscode
+.idea
+
+# Test artifacts
+htmlcov
+.coverage
+.pytest_cache
+playwright-report
+test-results
+
+# Misc
+*.log
+.DS_Store
+Thumbs.db

poly_hammer_worker-0.3.8/.gitignore

@@ -0,0 +1,19 @@
+node_modules/
+/test-results/
+/playwright-report/
+/blob-report/
+/playwright/.cache/
+scratches
+reports
+backend/reports
+.terraform
+.terraform-*
+*.tfvars
+infra/terraform/keys
+.vite/
+secret.env
+__pycache__/
+.hatchet/
+
+# Prebuilt wheels staged into the image build contexts by CI.
+workers/self-hosted/wheelhouse/

poly_hammer_worker-0.3.8/PKG-INFO

@@ -0,0 +1,14 @@
+Metadata-Version: 2.5
+Name: poly-hammer-worker
+Version: 0.3.8
+Summary: Poly Hammer self-hosted GPU worker — runs AI inference inside Docker
+License-Expression: MIT
+Requires-Python: <4,>=3.10
+Requires-Dist: hatchet-sdk>=1.32.0
+Requires-Dist: hf-transfer>=0.1.9
+Requires-Dist: httpx>=0.27
+Requires-Dist: modal>=1.3.2
+Requires-Dist: psutil>=5.9
+Requires-Dist: rich>=13.0
+Requires-Dist: sentry-sdk>=2.20
+Requires-Dist: typer>=0.27.1

poly_hammer_worker-0.3.8/README.md

@@ -0,0 +1,191 @@
+# Self-Hosted Worker Images
+
+Docker images for running Poly Hammer AI inference workers on your own hardware.
+Model weights are volume-mounted at runtime via named Docker volumes.
+
+> **Most people should use the `ph` CLI instead of the commands in this file.**
+> It picks the right CUDA build for your GPU and driver, manages tokens and
+> volumes, and starts the worker for you:
+>
+> ```bash
+> ph config
+> ph start
+> ```
+>
+> See [`workers/cli/README.md`](../cli/README.md). The rest of this document
+> covers building images locally and driving Docker by hand.
+
+## CUDA build variants
+
+Each GPU image is published in several CUDA builds. The portal's image catalog
+(`GET /v1/workers/images`) is the source of truth, and `ph` selects a build
+from it using your GPU's compute capability and driver version.
+
+| Image | Variant | CUDA / torch | Min driver | Compute capability |
+|-------|---------|--------------|-----------|--------------------|
+| `ph-hy-motion` | `cu121` | 12.1 / 2.5.1 | 530 | sm_75–sm_90 |
+| `ph-hy-motion` | **`cu128`** (default) | 12.8 / 2.7.1 | 570 | sm_75–**sm_120** |
+| `ph-hy-motion` | `cu130` | 13.0 / 2.9.0 | 580 | sm_75–sm_120 |
+| `ph-kimodo` | `cu126` | 12.6 / 2.5 | 525 | sm_75–sm_90 |
+| `ph-kimodo` | **`cu128`** (default) | 12.8 / 2.6 | 570 | sm_75–**sm_120** |
+| `ph-kimodo` | `cu130` | 13.0 / 2.9 | 580 | sm_75–sm_120 |
+
+The default variant is published as the bare tag (`:latest` / `:staging`) and
+also as an explicit `:latest-cu128`. Other variants use `:latest-<variant>`.
+
+Picking the wrong build is not a hard failure, which is what makes it
+dangerous: running Blackwell (RTX 50xx, sm_120) on a pre-12.8 build makes CUDA
+JIT-compile every kernel from PTX, costing minutes on each cold container.
+Prefer `ph`, which will not select such a build.
+
+CUDA 13 requires a 580+ driver, while every CUDA 12.x build works on any
+driver at or above its own floor thanks to minor-version compatibility.
+
+## Quick Start
+
+```bash
+# From workers/self-hosted/:
+docker compose build
+
+# Or build a single image / variant:
+docker compose build hy-motion
+docker compose build kimodo
+docker compose build kimodo-cu126
+docker compose build ph-hello-motion
+```
+
+This produces locally-tagged images:
+
+| Service | Image Tag | GPU Required |
+|---------|-----------|-------------|
+| HY-Motion | `ghcr.io/poly-hammer/ph-hy-motion:latest` | Yes |
+| HY-Motion (cu121) | `ghcr.io/poly-hammer/ph-hy-motion:latest-cu121` | Yes |
+| HY-Motion (cu130) | `ghcr.io/poly-hammer/ph-hy-motion:latest-cu130` | Yes |
+| Kimodo | `ghcr.io/poly-hammer/ph-kimodo:latest` | Yes |
+| Kimodo (cu126) | `ghcr.io/poly-hammer/ph-kimodo:latest-cu126` | Yes |
+| Kimodo (cu130) | `ghcr.io/poly-hammer/ph-kimodo:latest-cu130` | Yes |
+| Hello Motion | `ghcr.io/poly-hammer/ph-hello-motion:latest` | No |
+
+
+## Crash reports
+
+The worker reports crashes to Sentry unless told not to. `ph start` stamps
+`PH_TELEMETRY=0` or `=1` onto the container from the machine's `telemetry`
+setting, so declining in `ph config` silences the agent inside the image too.
+Driving Docker by hand, pass it yourself:
+
+```bash
+docker run --gpus all -e PH_TELEMETRY=0 ghcr.io/poly-hammer/ph-kimodo:latest
+```
+
+`PH_NO_TELEMETRY=1` does the same. Reports carry the exception, the package
+version, and the GPU model — never the machine name, an IP address, a token, or
+a prompt.
+
+
+## Hardware Requirements
+
+### HY-Motion (`ph-hy-motion`)
+
+| Resource | Minimum |
+|----------|---------|
+| GPU VRAM | 24 GB (Lite model) / 40 GB (full model) |
+| System RAM | 16–32 GB |
+| Disk (image) | ~40 GB |
+| Disk (build) | ~80 GB (temporary build layers) |
+
+### Hello Motion (`ph-hello-motion`)
+
+No GPU required. Uses a bundled motion.npz fixture to simulate the full
+inference pipeline. Useful for end-to-end testing of the worker system.
+
+### Kimodo (`ph-kimodo`)
+
+| Resource | Full GPU | CPU Text Encoder |
+|----------|----------|------------------|
+| GPU VRAM | 17 GB | 3 GB |
+| System RAM | 32 GB | 48 GB |
+| CUDA driver API | 12.6+ | 12.6+ |
+
+Kimodo is tested upstream on recent NVIDIA GPUs including RTX 3090/4090,
+A100, L4, and L40S. Set `TEXT_ENCODER_DEVICE=cpu` to use the lower-VRAM
+profile. On RTX 50xx use the default `cu128` build — see
+[CUDA build variants](#cuda-build-variants).
+
+## License Notices
+
+**You must review and accept the following licenses before building these images.**
+
+### HY-Motion (`ph-hy-motion`)
+
+- **Tencent Hunyuan Community License**: The HY-Motion model weights are released by Tencent under the [Tencent Hunyuan Community License Agreement](https://github.com/Tencent/HunyuanVideo/blob/main/LICENSE.txt). This license includes territory restrictions and usage conditions. Review it carefully.
+- **CLIP (OpenAI)**: The CLIP text encoder is released under the [MIT License](https://github.com/openai/CLIP/blob/main/LICENSE).
+- **Qwen3 (Alibaba)**: The Qwen3 text encoder is released under the [Apache 2.0 License](https://huggingface.co/Qwen/Qwen3-8B/blob/main/LICENSE).
+
+### Kimodo (`ph-kimodo`)
+
+- **Kimodo-SOMA-RP-v1.1**: NVIDIA Open Model License. The model is pinned to Hugging Face revision `6c9233af1180b8151e3c4703477104af5dce9dd5`.
+- **Kimodo code and MotionCorrection**: Apache 2.0, pinned to Kimodo commit `1aece8c124d73d255ceff5086d983b844c9f4e94`.
+- **LLM2Vec code/adapters**: MIT.
+- **Meta Llama 3 8B Instruct**: Meta Llama 3 Community License and Acceptable Use Policy. The model is manually gated on Hugging Face. Accept its terms before downloading and include the required notices in customer-facing deployments.
+
+Kimodo production deployment remains subject to legal approval of the complete
+model/encoder stack and required attribution. No Bones Rigplay or BONES-SEED
+dataset files are distributed by this worker.
+
+## Downloading Weights
+
+Weights are not baked into the image — download them once into a named volume:
+
+```bash
+docker volume create hy-motion-weights
+
+# Allocate a TTY (-it) so the live progress panel renders:
+docker run --rm -it --gpus all \
+    -v hy-motion-weights:/weights \
+    ph-hy-motion download-weights --job hy-motion --all
+```
+
+Expect roughly **50 GB** for `--all` and 20–60 minutes on a typical
+residential connection (`hf-transfer` is enabled by default).
+
+Kimodo requires the complete checkpoint and text-encoder closure. Its Meta
+Llama base model is gated, so pass an approved token:
+
+```bash
+docker volume create kimodo-weights
+
+docker run --rm -it --gpus all \
+    -v kimodo-weights:/weights \
+    -e HF_TOKEN \
+    ghcr.io/poly-hammer/ph-kimodo:latest \
+    download-weights --job kimodo --all
+
+docker run --rm --gpus all \
+    -v kimodo-weights:/weights \
+    -v "${PWD}/kimodo-output:/output" \
+    ghcr.io/poly-hammer/ph-kimodo:latest \
+    test-job --output-dir /output
+```
+
+The download command pins and verifies the Kimodo checkpoint, both LLM2Vec
+repositories, and `meta-llama/Meta-Llama-3-8B-Instruct`. Inference then runs
+with Hugging Face offline mode enabled so missing assets fail before a user job.
+
+### Helpful flags
+
+| Flag | Purpose |
+|------|---------|
+| `--hf-token <token>` | Authenticate with HuggingFace. Lifts anonymous rate limits and unlocks gated repos. Also reads `HF_TOKEN` / `HUGGING_FACE_HUB_TOKEN` env vars. |
+| `--no-progress` | Disable the live panel (use under `docker run -d` or in CI where there is no TTY). |
+| `--dry-run` | Print the download plan and exit without contacting HuggingFace. |
+| `--retries N` | Retry attempts per step on transient failures (default: 3, exponential backoff). |
+| `-v` / `--verbose` | Enable DEBUG logging from `ph-worker` and `huggingface_hub`. |
+| `-q` / `--quiet` | Suppress INFO logs; only show warnings, errors, and the progress panel. |
+
+### Why doesn't it look like anything is happening?
+
+If you ran `docker run` without `-t` (no TTY), Rich's live panel can't draw
+and you'll only see the periodic INFO log lines from `huggingface_hub`
+(e.g. `Fetching N files`). Pass `-it` for the panel, or `--no-progress` to
+opt into log-only mode.

poly_hammer_worker-0.3.8/compose.yml

@@ -0,0 +1,165 @@
+# =============================================================================
+# Self-Hosted Worker Images — Local Build
+#
+# Each image IS a Hatchet worker. Run directly:
+#   docker run --gpus all ph-hy-motion start --token <WORKER_TOKEN>
+#
+# Build all images (from workers/self-hosted/):
+#   docker compose build
+#
+# Build a single image:
+#   docker compose --env-file secret.env build hy-motion
+#   docker compose build ph-hello-motion
+#
+# Build a specific CUDA variant:
+#   docker compose build hy-motion-cu121
+#
+# Variant tags mirror the WORKER_IMAGES matrix in
+# backend/app/core/config/services.py — keep the two in sync.
+#
+# See README.md for license notices and hardware requirements.
+# =============================================================================
+
+services:
+  # ---------------------------------------------------------------------------
+  # HY-Motion — cu128 DEFAULT (CUDA 12.8 / torch 2.7.1, driver >=570)
+  # sm_75..sm_120 — native Blackwell (RTX 50xx) support.
+  # ---------------------------------------------------------------------------
+  hy-motion:
+    image: ghcr.io/poly-hammer/ph-hy-motion:latest
+    build:
+      context: .
+      dockerfile: poly_hammer_worker/jobs/hy_motion/Dockerfile
+    volumes:
+      - hy-motion-weights:/weights
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+  # ---------------------------------------------------------------------------
+  # HY-Motion — cu121 (CUDA 12.1 / torch 2.5.1, driver >=530)
+  # sm_75..sm_90 — for older drivers. No RTX 50xx support.
+  # ---------------------------------------------------------------------------
+  hy-motion-cu121:
+    image: ghcr.io/poly-hammer/ph-hy-motion:latest-cu121
+    build:
+      context: .
+      dockerfile: poly_hammer_worker/jobs/hy_motion/Dockerfile
+      args:
+        CUDA_TAG: 12.1.1-cudnn8-devel-ubuntu22.04
+        TORCH_INDEX_URL: https://download.pytorch.org/whl/cu121
+        TORCH_VERSION: "2.5.1"
+        TORCHVISION_VERSION: "0.20.1"
+    volumes:
+      - hy-motion-weights:/weights
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+  # ---------------------------------------------------------------------------
+  # HY-Motion — cu130 (CUDA 13.0 / torch 2.9.0, driver >=580)
+  # ---------------------------------------------------------------------------
+  hy-motion-cu130:
+    image: ghcr.io/poly-hammer/ph-hy-motion:latest-cu130
+    build:
+      context: .
+      dockerfile: poly_hammer_worker/jobs/hy_motion/Dockerfile
+      args:
+        CUDA_TAG: 13.0.1-cudnn-devel-ubuntu24.04
+        TORCH_INDEX_URL: https://download.pytorch.org/whl/cu130
+        TORCH_VERSION: "2.9.0"
+        TORCHVISION_VERSION: "0.24.0"
+    volumes:
+      - hy-motion-weights:/weights
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+  # ---------------------------------------------------------------------------
+  # Kimodo — cu128 DEFAULT (NGC 25.02 / CUDA 12.8 / torch 2.7, driver >=570)
+  # sm_75..sm_120 — native Blackwell (RTX 50xx) support.
+  # ---------------------------------------------------------------------------
+  kimodo:
+    image: ghcr.io/poly-hammer/ph-kimodo:latest
+    build:
+      context: .
+      dockerfile: poly_hammer_worker/jobs/kimodo/Dockerfile
+    volumes:
+      - kimodo-weights:/weights
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+  # ---------------------------------------------------------------------------
+  # Kimodo — cu126 (CUDA 12.6 / torch 2.6.0, driver >=525)
+  # sm_75..sm_90 — for older drivers. No RTX 50xx support.
+  # Plain CUDA base, not NGC: NGC's 12.6 images ship a pre-release torch that
+  # Kimodo's pinned transformers cannot import.
+  # ---------------------------------------------------------------------------
+  kimodo-cu126:
+    image: ghcr.io/poly-hammer/ph-kimodo:latest-cu126
+    build:
+      context: .
+      dockerfile: poly_hammer_worker/jobs/kimodo/Dockerfile.cuda
+      args:
+        CUDA_TAG: 12.6.3-cudnn-devel-ubuntu24.04
+        TORCH_INDEX_URL: https://download.pytorch.org/whl/cu126
+        TORCH_VERSION: "2.6.0"
+    volumes:
+      - kimodo-weights:/weights
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+  # ---------------------------------------------------------------------------
+  # Kimodo — cu130 (NGC 25.10 / CUDA 13.0 / torch 2.9, driver >=580)
+  # ---------------------------------------------------------------------------
+  kimodo-cu130:
+    image: ghcr.io/poly-hammer/ph-kimodo:latest-cu130
+    build:
+      context: .
+      dockerfile: poly_hammer_worker/jobs/kimodo/Dockerfile
+      args:
+        NGC_TAG: 25.10-py3
+    volumes:
+      - kimodo-weights:/weights
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+  # ---------------------------------------------------------------------------
+  # Hello Motion — Debug Image (no GPU)
+  # ---------------------------------------------------------------------------
+  ph-hello-motion:
+    image: ghcr.io/poly-hammer/ph-hello-motion:latest
+    build:
+      context: .
+      dockerfile: poly_hammer_worker/jobs/hello_motion/Dockerfile
+
+volumes:
+  hy-motion-weights:
+  kimodo-weights: