Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

DeepseekV4: Add fused hyper-connection ops (#25585)

Browse Source 
* dsv4 hc-ops

* add missing files;

* add cparams

* update rpc version

* address review comments

* address review comments
This commit is contained in:
Aman Gupta
2026-07-17 00:33:33 +08:00
committed by GitHub
parent b2dd28a3b6
commit 0dc74e332e
16 changed files with 1062 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-context.cpp
+31
View File
@@ -61,6 +61,24 @@ static const llm_fused_op_probe llm_fused_op_lid_probe = {
/*.n_tokens_per_seq =*/ 1,
};
static const llm_fused_op_probe llm_fused_op_dsv4_hc_pre_probe = {
/*.op =*/ LLM_FUSED_OP_DSV4_HC_PRE,
/*.name =*/ "fused DeepSeek V4 HC pre",
/*.n_tokens_per_seq =*/ 1,
};
static const llm_fused_op_probe llm_fused_op_dsv4_hc_comb_probe = {
/*.op =*/ LLM_FUSED_OP_DSV4_HC_COMB,
/*.name =*/ "fused DeepSeek V4 HC comb",
/*.n_tokens_per_seq =*/ 1,
};
static const llm_fused_op_probe llm_fused_op_dsv4_hc_post_probe = {
/*.op =*/ LLM_FUSED_OP_DSV4_HC_POST,
/*.name =*/ "fused DeepSeek V4 HC post",
/*.n_tokens_per_seq =*/ 1,
};
llama_context::llama_context(
const llama_model & model,
llama_context_params params) :
@@ -235,6 +253,11 @@ llama_context::llama_context(
cparams.fused_lid = true;
cparams.auto_flid = true;
cparams.fused_dsv4_hc_pre = true;
cparams.fused_dsv4_hc_comb = true;
cparams.fused_dsv4_hc_post = true;
cparams.auto_fhc = true;
// with causal attention, the batch size is limited by the context size
cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
@@ -537,6 +560,14 @@ void llama_context::resolve_fused_ops(const llama_memory_context_i * mctx, uint3
resolve(llm_fused_op_lid_probe, cparams.fused_lid);
cparams.auto_flid = false;
}
if (cparams.auto_fhc) {
LLAMA_LOG_INFO("%s: resolving fused DeepSeek V4 HC support:\n", func);
resolve(llm_fused_op_dsv4_hc_pre_probe, cparams.fused_dsv4_hc_pre);
resolve(llm_fused_op_dsv4_hc_comb_probe, cparams.fused_dsv4_hc_comb);
resolve(llm_fused_op_dsv4_hc_post_probe, cparams.fused_dsv4_hc_post);
cparams.auto_fhc = false;
}
}
void llama_context::sched_reserve() {