Repository navigation
Release v3.0.0.1
Model Compiler 3.0.0.1
Model Compiler 3.0.0.1 fixes the Qwen3.5 compilation performance regression and updates the 3.0 release with LLiMa compiler 0.6.1, Whisper compilation fixes, and newer frontend and compiler components.
Qwen3.5 compilation performance
Fixes a compiler regression that caused compiled Qwen3.5 models to run at lower tokens per second (TPS). Recompile Qwen3.5 with this release to adopt the fix; existing compiled artifacts retain the behavior of the compiler that produced them.
Whisper compilation fixes
The bundled LLiMa 0.6.1 compiler includes two fixes:
- GELU activation handling: newly compiled Whisper graphs use the tanh-based GELU implementation for source GELU activations. Compiler reference tests found lower numerical error.
- Decoder-cache input order: generated cache graphs use the runtime's expected
query, keys, mask, valuesorder. The incorrect order could corrupt transcription after the first word.
Recompile Whisper models to adopt the activation change. For packages affected by the cache-order bug, regenerate and replace the decoder-cache ELF; the encoder, initialization, and decoder pre/post ELFs can be retained. Existing working packages already use the expected order. Compiler graph checks passed, but full transcription correctness on hardware has not been established for every affected model.
See the upstream fixes: GELU handling and decoder-cache input order.
Component updates
| Component | 3.0.0 | 3.0.0.1 |
|---|---|---|
LLiMa compiler (sima_lmm[sdk]) |
0.6.0 | 0.6.1 |
sima-frontend |
3.0.0.dev0+master.402 |
3.0.0.dev0+master.403 |
sima-mlc |
3.0.0.dev0+master.193 |
3.0.0.dev0+master.194 |
The LLiMa compiler artifact is pinned to immutable ref v0.6.1:1bba98848cca. Other component pins remain unchanged from 3.0.0.
Build validation
Release preparation succeeded and verified the requested LLiMa 0.6.1 wheel before creating v3.0.0.1.
The tagged package build has passed packaging for amd64 and arm64. Package installation tests are still running.