ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 03:11:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

qwen3next: keep fused delta on safe path and remove PR artifacts

Browse Source
This commit is contained in:
yurko
2026-02-08 19:48:15 -08:00
parent 69529d3f49
commit 48e0e351ce
8 changed files with 13 additions and 1528 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
common/common.cpp
View File

@@ -1250,8 +1250,8 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
if (arg == "--qwen3next-fused-delta") {
CHECK_ARG
params.qwen3next_fused_delta = std::stoi(argv[i]);
if (params.qwen3next_fused_delta < 0 || params.qwen3next_fused_delta > 2) {
fprintf(stderr, "error: Invalid value for --qwen3next-fused-delta: %d (must be 0, 1, or 2)\n",
if (params.qwen3next_fused_delta < 0 || params.qwen3next_fused_delta > 1) {
fprintf(stderr, "error: Invalid value for --qwen3next-fused-delta: %d (must be 0 or 1)\n",
params.qwen3next_fused_delta);
invalid_param = true;
}
@@ -2181,7 +2181,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
options.push_back({ "*", " --chunks N", "max number of chunks to process (default: %d, -1 = all)", params.n_chunks });
options.push_back({ "*", "-no-fa, --no-flash-attn", "disable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled" });
options.push_back({ "*", "-fa, --flash-attn (auto|on|off|0|1)", "set Flash Attention (default: %s)", params.flash_attn ? "on" : "off" });
options.push_back({ "*", " --qwen3next-fused-delta {0,1,2}",
options.push_back({ "*", " --qwen3next-fused-delta {0,1}",
"force LLAMA_QWEN3NEXT_FUSED_DELTA mode for Qwen3Next (default: env/model default)" });
options.push_back({ "*", "-mla, --mla-use", "enable MLA (default: %d)", params.mla_attn });
options.push_back({ "*", "-amb, --attention-max-batch", "max batch size for attention computations (default: %d)", params.attn_max_batch});

2
common/common.h
View File

@@ -259,7 +259,7 @@ struct gpt_params {
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
bool cont_batching = true; // insert new sequences for decoding on-the-fly
bool flash_attn = true; // flash attention
int qwen3next_fused_delta = -1; // -1 keep env/default, otherwise force LLAMA_QWEN3NEXT_FUSED_DELTA={0,1,2}
int qwen3next_fused_delta = -1; // -1 keep env/default, otherwise force LLAMA_QWEN3NEXT_FUSED_DELTA={0,1}
int mla_attn = 3; // MLA 0: standard, 1: MLA with K and V^T cache, 2: MLA with just K cache, 3: the best of both worlds
int attn_max_batch = 0; // Max batch size to use when computing attention (only applicable if flash_attn = false)
bool fused_moe_up_gate = true; // fused up*unary(gate) op for MoE models