mlboydaisuke
/

Qwen3.8-27B-CoreAI

Image-Text-to-Text

speculative-decoding

Model card Files Files and versions

Qwen3.8-27B-CoreAI / gpu-pipelined

107 GB

Ctrl+K

Ctrl+K

1 contributor

History: 9 commits

mlboydaisuke's picture

qwen3_8_27b_verify_s9_int4lin_d4_hpost: int4 S=9 verify + hidden out for the MTP drafter (29.0 tok/s code, lossless)

aaccef9 verified 2 months ago

qwen3_8_27b_decode_int4lin
qwen3_8_27b_decode_int4lin: int4 text decoder (22.2 tok/s M4 Max, gate 15/16) 2 months ago
qwen3_8_27b_decode_int8hu_block32_sym
Qwen3.8-27B: qwen3_8_27b_decode_int8hu_block32_sym 2 months ago
qwen3_8_27b_mtp_s1_int8hu_block32_sym
qwen3_8_27b_mtp_s1_int8hu_block32_sym: MTP head as S=1 stateful drafter (int8) 2 months ago
qwen3_8_27b_mtp_s9_int8hu_block32_sym
qwen3_8_27b_mtp_s9_int8hu_block32_sym: MTP S=9 replay graph (shared-KV sibling) 2 months ago
qwen3_8_27b_verify_s9_int4lin_d4_hpost
qwen3_8_27b_verify_s9_int4lin_d4_hpost: int4 S=9 verify + hidden out for the MTP drafter (29.0 tok/s code, lossless) 2 months ago
qwen3_8_27b_vision_fp16
Qwen3.8-27B: qwen3_8_27b_vision_fp16 2 months ago
qwen3_8_27b_vl_decode_int8hu_block32_sym_pf16
Qwen3.8-27B: VL decoder pf16 (pf32 chunk overflows fp16 content-dependently) 2 months ago