[EAGLE] Refactor code for page size > 1 & more simplifications (#7213)

This commit is contained in:
Lianmin Zheng
2025-06-16 03:04:29 -07:00
committed by GitHub
parent 21615cc3fe
commit b1286a116a
8 changed files with 647 additions and 156 deletions

View File

@@ -1049,14 +1049,13 @@ class FlashInferMultiStepDraftBackend:
kv_indices_buffer,
self.kv_indptr,
forward_batch.positions,
num_seqs,
self.topk,
self.pool_len,
kv_indices_buffer.shape[1],
self.kv_indptr.shape[1],
next_power_of_2(num_seqs),
next_power_of_2(self.speculative_num_steps),
next_power_of_2(bs),
self.page_size,
)
assert forward_batch.spec_info is not None