feat(escha): serve Qwen3.6-35B-A3B-Escha-W2 natively (2-bit eschamoe + int8), lossless AR ~58 tok/s - #242

Open

davidtai wants to merge 7 commits into

youssofal:mainfrom

davidtai:feat/escha-w2-native

Commits on Aug 5, 2026

feat(escha): serve EschaLabs Qwen3.6-35B-A3B-Escha-W2 natively with MTP spec-decode

davidtai
and
claude
committed
039b7a2
perf(escha): eschamoe bf16 compute default + int8 fused-decode/GEMM-prefill split

davidtai
and
claude
committed
d1b3ec7

Commits on Aug 6, 2026

perf(escha): bf16 scatter buffer in the grouped prefill MoE (was f32, 2x activation bytes for nothing)

davidtai
and
claude
committed
232502e
perf(escha): mx.compile the batch=1 eschamoe decode chain (ESCHA_COMPILE)

davidtai
and
claude
committed
61423e9
Add opt-in async-eval pipeline to single-stream AR serving decode

davidtai
and
claude
committed
2f26802
docs(escha): AR-focused serving evidence (perf/mem + MTP status)

davidtai
and
claude
committed
31aeb68
docs(escha): full 1024-128k prefill/decode/memory benchmark; keep grouped prefill

davidtai
and
claude
committed
614229d