Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions docs/features/nixl_connector_compatibility.md
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,7 @@ th:not(:first-child) {
| Dense Transformers | ✅ | ✅<sup>1</sup> | ✅ | ✅<sup>2</sup> | ✅ | ✅ | 🟠<sup>3</sup> |
| MLA (e.g. DeepSeek-V2/V3) | ✅ | ✅<sup>1</sup> | 🟠<sup>4</sup> | ✅<sup>2</sup> | ✅ | ✅ | 🟠<sup>3</sup> |
| Sparse MLA (e.g. DeepSeek-V3.2) | ✅ | ✅<sup>1</sup> | 🟠<sup>4</sup> | ✅<sup>2</sup> | ✅ | ✅ | 🟠<sup>3</sup> |
| Hybrid SSM / Mamba | ✅ | ❔ | 🚧<sup>5</sup> | ❌ | ✅ | ✅ | ❌<sup>6</sup> |
| Hybrid SSM / Mamba | ✅ | ❔ | 🟠<sup>5</sup> | ❌ | ✅ | ✅ | ❌<sup>6</sup> |
| MoE | ✅ | ✅<sup>1</sup> | ✅ | ✅<sup>2</sup> | ✅ | ✅ | 🟠<sup>3</sup> |
| Multimodal | ❔ | ❔ | ❔ | ❔ | ❔ | ❔ | ❔ |
| Encoder-Decoder | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
Expand All @@ -67,7 +67,7 @@ what may differ.

<sup>4</sup> MLA KV cache is replicated across TP workers, so heterogeneous TP works but there is no head-splitting. When P TP > D TP, only a single read is executed (redundant ranks are skipped). D TP > P TP also works.

<sup>5</sup> Hybrid SSM (Mamba) models require **homogeneous TP** (`P TP == D TP`). Heterogeneous TP is not yet supported for Mamba layers.
<sup>5</sup> Hybrid SSM/Mamba models (including GDN) support heterogeneous TP in both directions, provided one TP size is an integer multiple of the other.

<sup>6</sup> HMA (required by hybrid models) does not support different remote block sizes.

Expand Down
Loading