[PyTorch] DeepSeekV3Layer: full MoE transformer layer (MLA + DeepSeek MoE) - #36

Draft

pggPL wants to merge 12 commits into

mainfrom

deepseek_v3_layer

Commits on Aug 18, 2026

[PyTorch] Add DeepSeekV3Layer skeleton (MLA + MoE)

pggPL
and
claude
committed
45ed767
Move DeepSeekV3 skeleton to models/deepseek_v3 subpackage

pggPL
and
claude
committed
c306c6f
Add DeepSeekV3 layer entries to PyTorch API docs

pggPL
and
claude
committed
f73d04e
[PyTorch] Implement DeepSeekV3Layer: MLA + MoE from TE building blocks

pggPL
and
claude
committed
09f28a9
Add distributed EP test for DeepSeekV3 MoE/layer

pggPL
and
claude
committed
e23100b
Fix EP wgrad test collective + zero EP recv/grad buffers

pggPL
and
claude
committed
4c6e1e8
Use fused MLA RoPE kernels in MultiLatentAttention

pggPL
and
claude
committed
aa17c37

Commits on Aug 21, 2026

Add HF transformers numeric reference test for DeepSeekV3Layer

pggPL
and
claude
committed
c713af7
Docstring cleanups for lint and docs build

pggPL
and
claude
committed
88028a9
Move model-specific layers to a dedicated docs page

pggPL
and
claude
committed
5f68c9b
Drop HF-transformers comparison test from the repo

pggPL
and
claude
committed
9db495a
Docs: reduce models page to a plain API listing

pggPL
and
claude
committed
4883b17