Support for Deepseek V4 flash

[sorasoras]

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
https://github.com/deepseek-ai/DeepGEMM
280B/A13
Support for inference on the SOTA model at this size with 1M context

[mratsim]

Blocked by MLA just like DeepSeek, GLM-4.7-Flash and Kimi-K2.x:

• [Support request] zai-org/GLM-4.7-Flash #142
• Feedback wanted: native MLA baseline with optional flash-attn acceleration #175

And then DeepSeek V4 requires DSA and NSA/Lightning Indexer on top. And you need to implement mHC as well.

It's not even in vLLM / SGLang for non-H100 / B200 hardware

[Ph0rk0z]

V4 (and I think v3.1 as well) has filtered everyone.

[Knightcodin]

Perhaps the blocker is MLA + Sequence-Dimension Compression?