[NPU] Support DFlash speculative decoding for MiMo-V2.5-Pro (mxfp4) (#37565)

This commit is contained in:
iridiumine
2026-09-12 11:59:23 +08:00
committed by GitHub
parent 9c365e97b1
commit a984c78330
15 changed files with 741 additions and 80 deletions
@@ -149,6 +149,12 @@ META_LLAMA_3_1_8B_INSTRUCT = os.path.join(
)
MIMO_7B_RL_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "XiaomiMiMo/MiMo-7B-RL")
MIMO_V2_FLASH_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "XiaomiMiMo/MiMo-V2-Flash")
MIMO_V2_5_PRO_FP4_DFLASH_WEIGHTS_PATH = os.path.join(
MODEL_WEIGHTS_DIR, "XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash"
)
MIMO_V2_5_PRO_FP4_DFLASH_DRAFT_WEIGHTS_PATH = os.path.join(
MIMO_V2_5_PRO_FP4_DFLASH_WEIGHTS_PATH, "dflash"
)
MIMO_V2_5_W8A8_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "solinliu/MiMo-V2.5-W8A8")
MINICPM3_4B_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "OpenBMB/MiniCPM3-4B")
MISTRAL_7B_INSTRUCT_V0_2_WEIGHTS_PATH = os.path.join(