Support for DeepseekV32ForCausalLM with DeepSeek Sparse Attention (DSA) - #21149

Closed

fairydreaming wants to merge 80 commits into

ggml-org:masterfrom

fairydreaming:deepseek-dsa

Commits on Mar 12, 2026

model : Initial support for DeepseekV32ForCausalLM (for now with dense attention). Needs manual change of add_bos_token to true in tokenizer_config.json before conversion.
sszymczy
committed
a337ebd
model : added indexer q and k calculation in DeepseekV32ForCausalLM.
sszymczy
committed
e467684
ggml : add Hadamard transform GGML OP and implementation
sszymczy
committed
723f0ce

Commits on Mar 13, 2026

kv-cache : add cache for indexer keys (temporary solution)
sszymczy
committed
72b7214

Commits on Mar 14, 2026

Commits on Mar 15, 2026

Commits on Mar 16, 2026

Commits on Mar 22, 2026

model : replaced ggml_argsort_top_k with ggml_top_k in DeepSeek V3.2 indexer implementation since the former fails for large tensors even when using CCCL.
sszymczy
committed
e7aa89a

Commits on Mar 23, 2026

model : added comments in DeepSeek V3.2 lightning indexer implementation.
sszymczy
committed
1874ac9

Commits on Mar 24, 2026

Commits on Mar 25, 2026

Commits on Mar 31, 2026

Merge remote-tracking branch 'upstream/master' into deepseek-dsa
sszymczy
committed
57a8def

Commits on Apr 1, 2026

Commits on Apr 2, 2026

Commits on Apr 8, 2026

ggml : added new GGML_OP_LIGHTNING_INDEXER that merges ggml_mul_mat(), ggml_relu() and ggml_sum_rows() to save compute buffer memory.
sszymczy
committed
014e63c

Commits on Apr 12, 2026

ggml : support lightning indexer key quantization
sszymczy
committed
3d61d0b

Commits on Apr 13, 2026

Merge remote-tracking branch 'upstream/master' into deepseek-dsa
sszymczy
committed
ec083ce
chore : fix trailing whitespaces
sszymczy
committed
65c3557
ggml : bump RPC version
sszymczy
committed
5715c36
chore : silence Python Type-Check CI error
sszymczy
committed
e0fcb22
ggml : more assertions in GGML_OP_SCATTER since there is no broadcasting
sszymczy
committed
45294c5
model : corrected number of layers for 685B.A37B DeepseekV32ForCausalLM models
sszymczy
committed
66ec7be

Commits on Apr 14, 2026

Commits on Apr 15, 2026

Commits on Apr 20, 2026

chore : whitespaces
sszymczy
committed
9695fc8
Merge remote-tracking branch 'upstream/master' into deepseek-dsa
sszymczy
committed
938ef66
Merge remote-tracking branch 'upstream/master' into deepseek-dsa
sszymczy
committed
789ec5a

Commits on Apr 21, 2026

Merge remote-tracking branch 'upstream/master' into deepseek-dsa
sszymczy
committed
a51a7d4

Commits on Apr 24, 2026

Commits on Apr 25, 2026

ggml : pad strides of shared memory Q and K buffers to reduce bank conflicts in wmma lightning indexer implementation.
sszymczy
committed
c018008

Commits on Apr 28, 2026

ggml : new ggml_flash_attn_ext_add_top_k() function for adding top_k information to flash attention OP; added optimization case in CUDA fattn-mma kernel to use top_k for faster token generation.
sszymczy
committed
65ab575

Commits on May 1, 2026

Commits on May 3, 2026

ggml : add missing threads sync, replace magic numbers with proper constants in lightning indexer kernel
sszymczy
committed
59769fe

Commits on May 4, 2026

Commits on May 5, 2026

Commits on May 6, 2026