08-models_deepseek_v4_1_compressor.py.diff 438 B

1234567891011
  1. --- baked-ours/models/deepseek_v4_1/compressor.py 2026-09-10 19:56:14.000000000 +0200
  2. +++ baked/models/deepseek_v4_1/compressor.py 2026-09-11 10:13:47.000000000 +0200
  3. @@ -160,6 +160,8 @@
  4. head_size=self.state_dim,
  5. head_size_v=0,
  6. dtype=self.dtype,
  7. + # Every DCP rank compresses every token: the ring is replicated.
  8. + dcp_replicated=True,
  9. )
  10. def forward(self): ...