biondizzle 5003e756e2 WIP: cross-check 2 fix — block-aligned compressed RoPE positions + compress_rope_theta support ...

- CRITICAL BUG FIX: comp_pos was using LAST position of each block (((bi+1)*r-1))
  instead of FIRST position (bi*r). Off by r-1: 3 for CSA, 127 for HCA.
  vLLM uses (position // ratio) * ratio = block-aligned first position.
- Added compress_rope_theta config support (vLLM uses separate theta for compressed)
- Added comp_rope_cos/comp_rope_sin param to forward_layer (not yet wired through)

Only single_shot_inference.py changed — no kernel code touched.
Base commit: 572bdd2

2026-06-03 09:17:54 +00:00

..

deepseek_v4_ref/deepseek_v4

WIP: cross-check 2 fix — block-aligned compressed RoPE positions + compress_rope_theta support

2026-06-03 09:17:54 +00:00

CROSS_REFERENCE.md

auto: pre-test commit

2026-06-03 09:01:02 +00:00

deepseek_v4.zip

Test 2: fix topk tensor shape (flatten before iterating)

2026-06-03 08:47:32 +00:00