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