Feat/moe nsp blocking all models#1016

Open

divytrip3005 wants to merge 16 commits into

divytrip3005:feat/moe-nsp-blocking-all-models

Commits on Apr 30, 2026

Commits on May 13, 2026

Commits on May 14, 2026

fix(0415): fix: avoid unsupported prefill MoE reductions in subfunction export
vbaddi
committed

Commits on May 15, 2026

fix(0415): align prefill MoE chunk export with packed dispatch
vbaddi
committed

Commits on May 21, 2026

feat: NSP-blocked MoE prefill for GPT-OSS, Qwen3-MoE and GraniteMoE
divytrip
committed
fix: replace torch.clamp with torch.where for int32 chunk_valid_rows
divytrip
committed

Commits on May 26, 2026

feat: API-driven NSP blocking for Qwen3-VL-MoE, Qwen3-MoE, GPT-OSS, GraniteMoE
divytrip
committed

Commits on May 27, 2026

fix: add num_cores and moe_prefill_packed_chunk_size as explicit params to QEFFAutoModelForCausalLM.export()
divytrip
committed

Commits on Jun 1, 2026

fix: moe_prefill_num_nsp API param + GraniteMoE NSP blocking fixes
divytrip
committed