tt-skyreels

SkyReels-V2-DF-1.3B-540P (text-to-video) on Tenstorrent Blackhole via TTNN -- reuses tt-metal's WanTransformer3DModel wholesale (SkyReelsV2's transformer is weight-compatible with it) behind a small ASGI server, following the tt-dit-server pattern tt-animatediff established.

Runs on p300x2 (mesh QB2).

Packaged and published with tt-model-manager as a v6 thin bundle (schema 6) -- a pip/venv-based install, not a container image.

Quickstart

tt-model pull  episod/tt-skyreels --with-weights
tt-model serve episod/tt-skyreels

pull --with-weights downloads this bundle (a small venv-install recipe: install.sh/run.sh + two wheels -- ttnn, torch, the diffusers stack, and the base HTTP stack resolve from an index at install time) and the Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers weights, into your HF cache. serve starts the model's own HTTP server on port 20000 (or the next free port, if that one is busy); the first start compiles kernels for your device, which takes several minutes, and the server is ready when it logs Application startup complete.

tt-model pull episod/tt-skyreels --with-weights
tt-model serve episod/tt-skyreels
curl localhost:20000/v1/videos/generations \
  -H 'Content-Type: application/json' \
  -d '{"prompt": "a red bicycle leaning against a brick wall in soft morning light"}'

Source, a local Gradio UI, and the bring-up log (bugs found only by actually serving a generation, none of them in the reused TTNN model itself): tsingletaryTT/tt-skyreels.

Provenance

The served-path code is split into two small wheels built for this bundle, both in wheels/:

wheel what it is
skyreels-ttnn-0.1.0 this repo's served-path closure -- session.py, the ASGI app, pipeline_skyreels.py -- built from its own existing setup.py
tt_skyreels_models_closure-0.78.0 tt-metal's models/tt_dit (SkyReelsV2's transformer is weight-compatible with WanTransformer3DModel and reuses it wholesale) plus the two models/common modules it actually imports, pinned to tt-metal v0.78.0

kernel_patch/: ttnn 0.78.0's PyPI wheel omits three kernel source files (fabric_router_mux_extension.cpp, fabric_router_relay_extension.cpp, fabric_router_udm_mux_extension.cpp) that FabricConfig::FABRIC_1D needs to JIT-compile its mux/relay routing kernels -- present in the real tt-metal v0.78.0 source tree, absent from the wheel. Vendored here byte-for-byte from that source tree and made discoverable via TT_METAL_KERNEL_PATH (an existing tt_metal env var, set in run.sh) rather than anything specific to this model. See kernel_patch/README.md.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support