Commit History
3 days ago
-
README: 1.0.6 is backward compatible with 1.0.4 / 1.0.5 (how it was checked)
Verified signatureDocs only. States that the release is a drop-in upgrade and how it was checked: same model files, serve.sh and dependencies; every native operation of the previous wheels kept with the same signature; no Python API removed; greedy output compared byte-for-byte against the previous release on an RTX 4090.9969344 -
1.0.6: faster single-user decode (fused MoE routing + shared expert, fused attention glue)
Verified signatureRuntime-only point release; the model repo is unchanged and every launch command works as before. Decode (up to 8 tokens per step) runs the router, top-k and expert tables in one kernel and folds the shared expert into the expert scatter; the attention glue is two fused kernels (bit-identical). Knobs SGLANG_ESCHA_MOE_GLUE=0 / ESCHA_MOE_ROUTE_BUILD=0 / SGLANG_ESCHA_ATTN_FUSE=0 restore 1.0.5.5b15ee8 -
1.0.5: faster prefill (grouped MoE prefill kernel) and the latest escha kernels
Verified signatureRuntime-only point release; the model repo is unchanged and every launch command works as before. Prefill batches run the routed experts as a grouped tile GEMM (ESCHA_MOE_GROUPED_PREFILL=0 restores the 1.0.4 path); escha._C is rebuilt from the current kernels with the same op surface.292e93f
5 days ago
-
1.0.4: opt-in MTP speculative decoding (MTP=1) for the SGLang engine
Verified signatureRuntime-only point release; the model repo is unchanged. New default-off knobs MTP and HOTMAP; every existing setting launches the same server as 1.0.3. escha._C is byte-identical to 1.0.3.8ead6e3
-
README: fix stale torch==2.9.* pin in the Quickstart snippet
Verified signaturedb8754c
-
escha runtime 1.0.3 (qwen3moe) — stable-ABI wheel, torch 2.9 and later
Verified signature2128884
-
add MLX runtime
Verified signature38a8508
-
document the 32k single-stream option on 16GB cards
Verified signaturec7ec5a7
-
Escha runtime — SGLang wheel + ZML single-binary server, with verified per-GPU recipes
90cf55e