Port DeepSeek-v4 to TQ - #224

Merged

TheTom merged 16 commits into

TheTom:feature/turboquant-kv-cachefrom

Jul 30, 2026

Commits on Jul 18, 2026

ds4 : turboquant feature work
giveen
committed
2d18e89
Merge origin/feature/turboquant-kv-cache into ds4
giveen
committed
c94612e
ds4 : warp-parallel HC_COMB kernel + add test cases
giveen
committed
4f5cb54
cuda : PDL + __restrict__ fix for flash_attn_mask_to_KV_max
giveen
committed
95a7c30
ds4 : port upstream graph fixes and l_last backend assignment
giveen
committed
3ba685b
llama-quant : exclude ffn_gate_tid2eid from quantization
giveen
committed
097c369
ggml : add a set of functions for checking contiguity of inner tensor dimensions (#25650)

authored and
giveen
committed
9931856
ggml : relax contiguity requirements for quantized concat

sszymczy
authored and
giveen
committed
8b843d3
cuda : relax tensor contiguity requirements for quantized concat

sszymczy
authored and
giveen
committed
5d3033e
cuda : remove stale quantized-type assert from concat
giveen
committed
682a25d

Commits on Jul 30, 2026