Spaces:
Running on Zero
Running on Zero
Commit History
-
Patch prompt/target separator back to a space Running
Upstream 0.3.1 joins prompt text and target text with a trailing newline (prompt_text += "\n"). "\n" is under-represented in the training data and makes voice cloning behave unevenly, so restore a plain space separator, applied uniformly for every language. The patch wraps DotsTtsRuntime._process_prompt_text rather than reimplementing it (stays decoupled from upstream's language-tag handling) and self-checks the upstream newline convention at install time, falling back to no-op if it drifts. Env-toggle: DOTS_TTS_PROMPT_TEXT_SPACE_SEP (default on).1a1a698
-
AOTI-compile the LLM decode step on ZeroGPU
The LLM per-token decode step (LLMInference._static_inputs_embeds) previously ran eager because torch.export/AOTInductor cannot express StaticCache's in-place KV mutation (functionalization aborts). Compile it via a functional rewrite. - dots_tts_zerogpu/llm_step.py: FuncStaticLayer overrides only StaticLayer.update to be out-of-place (torch.index_copy); LlmDecodeStep wraps the forward as a pure (emb, kv, cumulative_length, cache_position) -> (hidden, new_kv) function so no input is mutated in place. install_llm_step patches _get_static_token_step to marshal the serving StaticCache through the compiled step and write KV back in eager Python, and routes the runner across the ZeroGPU fork via AOTI_REGISTRY. - Startup self-check compares compiled vs eager hidden (relative-norm + cosine); on mismatch/failure the LLM step silently stays eager. DOTS_TTS_AOTI_LLM_STEP=0 disables it; DOTS_TTS_AOTI_STRICT=1 raises instead of falling back. - aoti.py: mark targets uncompilable afte…4f4dba4 -
Fix Space sdk_version: 6.17.0 does not exist, use 6.17.3
db8a33c -
Deploy via pip-installed dots.tts + thin ZeroGPU AOTI patch
Stop vendoring the dots.tts source tree. The Space now installs the published package (dots.tts==0.3.1) and applies all ZeroGPU specifics through a small local patch layer, so upstream updates no longer require re-aligning forked model code. - add dots_tts_zerogpu/: intercepts upstream's single compile seam (inference_utils compile_module_forward / compile_bound_method) and routes it to spaces.aoti_compile with a cross-fork artifact registry; drives compilation through upstream run_warmup; forces the sdpa attention backend (flex relies on module-level torch.compile). - rewrite app.py ZeroGPU bootstrap around the patch layer. - requirements.txt: pin dots.tts==0.3.1 + spaces, drop vendored deps. - remove src/ and training configs; default model -> dots-studio/dots.tts-soar.1c7476a
-
support pinning HF model revision via DOTS_TTS_MODEL_REVISION env var
241b819 -
Revert "use reflect-pad extra patch on prompt audio instead of regenerate-and-drop"
a12fb8c
-
bump default num_steps to 16
f48822c -
use reflect-pad extra patch on prompt audio instead of regenerate-and-drop
332a4c1 -
switch default execution mode to generate
780bcb5 -
add voice cloning tips card to UI
3d3f84e
-
add constraint on both prompt text and text to generate
d47a173 -
update constraints for better performance
3c5d8a1
-
add some constraints for better performance
f2be9ce
-
set default generation time to 60s
927c02a -
copy from /tmp to persist bucket to avoid gradio error
21b176e -
refine Gradio placeholder translation
2e1b373 -
change generated dir to huggingface space mounted bucket
ba03053 -
add an import
50c5224 -
add inference code with AOTI support for hf space
804ee23 -
initial commit
Verified signature02617d2
