Feature Combination Matrix#
The general combinations below are subject to each model’s attention constraints. For DeepSeek-V4.1 CSA2, use these additional limits:
Feature |
CSA2 constraint |
|---|---|
Hardware / shape |
SM90, SM100 family or SM120/121; one KV head, head dimension 512, checkpoint-owned compression and source layout |
TP / EP / attention DP |
Supported by the model integration |
PP / CP / disabled layers |
Unsupported |
Overlap / CUDA graphs |
Supported with graph-owned metadata and serialized shared arenas |
Chunked prefill / KV reuse |
Supported; GLOBAL is REQUIRED, Encoder SWA may be OPTIONAL, bounded Decoder SWA is PRIVATE |
Speculative decoding |
Embedded DSpark linear verification; tree verification unsupported |
Disaggregated serving |
Role mappings and cache transfer have component coverage; end-to-end serving requires separate validation |
Bounded replay |
Approximate Decoder replay defaults on for eligible layouts; Encoder recovery is opt-in |
See CSA2 sparse attention for staging dtypes, replay environment variables, and accuracy-validation requirements.
Feature |
Overlap Scheduler |
CUDA Graph |
Tensor Parallelism |
Pipeline Parallelism |
Expert Parallelism |
Helix Parallelism |
Attention Data Parallelism |
Disaggregated Serving |
Chunked Prefill |
Speculative Decoding — Linear |
Speculative Decoding — Dynamic Trees |
Speculative Decoding — Legacy Path (NGram, user-provided) |
Torch Sampler |
KV Cache Reuse |
Sliding Window Attention |
Logits Post Processor |
Guided Decoding |
LoRA |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Overlap Scheduler |
— |
|||||||||||||||||
CUDA Graph |
Yes |
— |
||||||||||||||||
Tensor Parallelism |
Yes |
Yes |
— |
|||||||||||||||
Pipeline Parallelism |
Yes |
Yes |
Yes |
— |
||||||||||||||
Expert Parallelism |
Yes |
Yes |
Yes |
Yes |
— |
|||||||||||||
Helix Parallelism |
Untested |
Yes |
Yes |
Yes |
Yes |
— |
||||||||||||
Attention Data Parallelism |
Yes |
Yes |
Yes |
Yes |
Yes |
Known issues |
— |
|||||||||||
Disaggregated Serving |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
— |
||||||||||
Chunked Prefill |
Yes |
Yes |
Yes |
Untested |
Yes |
Yes |
Yes |
Yes |
— |
|||||||||
Speculative Decoding — Linear |
Yes |
Yes |
Yes |
No |
Yes |
No |
Yes |
Yes |
Yes |
— |
||||||||
Speculative Decoding — Dynamic Trees |
Yes |
Yes |
Yes |
No |
Yes |
No |
Yes |
Yes |
Yes |
No |
— |
|||||||
Speculative Decoding — Legacy Path (NGram, user-provided) |
Yes |
Yes |
Yes |
No |
Yes |
No |
Yes |
Yes |
Yes |
No |
No |
— |
||||||
Torch Sampler |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes (MTP dynamic tree: greedy only) |
Yes |
— |
|||||
KV Cache Reuse |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
— |
||||
Sliding Window Attention |
Yes |
Yes |
Yes |
Yes |
Yes |
Untested |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
— |
|||
Logits Post Processor |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
No |
Yes |
No |
No |
No |
Yes |
Yes |
Yes |
— |
||
Guided Decoding |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
Yes |
— |
|
LoRA |
Yes |
Yes |
Yes |
Yes |
Untested |
Untested |
Untested |
Untested |
Yes |
Untested |
Untested |
Untested |
Yes |
Yes |
Yes |
Yes |
Untested |
— |