Fix Mamba2 chunked-prefill / speculative decoding for Zamba2, Nemotron-H, Bamba, FalconH1 and GraniteMoeHybrid #46741
Fix Mamba2 mixer chunked-prefill/speculative decode for Zamba2 and Ne…
2c8986fe
Prepend the full cached conv_states in the Mamba2 chunked path
1aaaeb62
vasqu
commented
on 2026-06-22
Fix Mamba2 cuda_kernels_forward chunked-prefill for Zamba2 and Nemotr…
60a85f8f
Align Zamba2/Nemotron-H chunked-prefill tests with merged mamba2 test
3514e781
Propagate Mamba2 chunked-prefill cache fix to Bamba, FalconH1, Granit…
91b8f8ce
Sunt-ing
changed the title Fix Mamba2 mixer chunked-prefill / speculative decoding for Zamba2 and Nemotron-H Fix Mamba2 chunked-prefill / speculative decoding for Zamba2, Nemotron-H, Bamba, FalconH1 and GraniteMoeHybrid 99 days ago
vasqu
commented
on 2026-06-26
Extract cached conv/recurrent state into locals in the Mamba2 mixers
a0320f55
Add CPU variant for the nemotron_h chunked-prefill test
e351f092
Merge branch 'main' into c001-mamba2-chunked-prefill
14f547f6
rename a bit
7bd3cada
vasqu
approved these changes
on 2026-06-29
vasqu
merged
fdb6d318
into main 96 days ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub