RYDE-WORK/ktransformers
1
0
Fork 0
mirror of https://github.com/RYDE-WORK/ktransformers.git synced 2026-02-05 14:03:17 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

⚡ fix moe.cpp int overflow problem

Browse Source
This commit is contained in:
liam 2025-02-06 22:39:16 +08:00
parent ee24a27001
commit 3dca28d23b
2 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ktransformers/ktransformers_ext/operators/llamafile/moe.cpp
View File

@ -224,7 +224,7 @@ void MOE::forward_many(int qlen, int k, const uint64_t* expert_ids, const float*
int stride = QK_K; int stride = QK_K;
int nth = config_.intermediate_size / stride; int nth = config_.intermediate_size / stride;
backend->do_work_stealing_job(nth * config_.expert_num, nullptr, [&](int task_id) { backend->do_work_stealing_job(nth * config_.expert_num, nullptr, [&](int task_id) {
int expert_idx = task_id / nth; uint64_t expert_idx = task_id / nth;
int ith = task_id % nth; int ith = task_id % nth;
void* gate_input_ptr = m_local_gate_input_ptr_[expert_idx]; void* gate_input_ptr = m_local_gate_input_ptr_[expert_idx];
void* gate_proj_ptr = (uint8_t*)gate_proj_ + (expert_idx * config_.intermediate_size + ith * stride) * config_.hidden_size * ggml_type_size(config_.gate_type) / ggml_blck_size(config_.gate_type); void* gate_proj_ptr = (uint8_t*)gate_proj_ + (expert_idx * config_.intermediate_size + ith * stride) * config_.hidden_size * ggml_type_size(config_.gate_type) / ggml_blck_size(config_.gate_type);
@ -246,7 +246,7 @@ void MOE::forward_many(int qlen, int k, const uint64_t* expert_ids, const float*
stride = QK_K; stride = QK_K;
nth = config_.hidden_size / stride; nth = config_.hidden_size / stride;
backend->do_work_stealing_job(nth * config_.expert_num, nullptr, [&](int task_id) { backend->do_work_stealing_job(nth * config_.expert_num, nullptr, [&](int task_id) {
int expert_idx = task_id / nth; uint64_t expert_idx = task_id / nth;
int ith = task_id % nth; int ith = task_id % nth;
void* down_input_ptr = m_local_down_input_ptr_[expert_idx]; void* down_input_ptr = m_local_down_input_ptr_[expert_idx];
void* down_proj_ptr = (uint8_t*)down_proj_ + (expert_idx * config_.hidden_size + ith * stride) * config_.intermediate_size * ggml_type_size(config_.down_type) / ggml_blck_size(config_.down_type); void* down_proj_ptr = (uint8_t*)down_proj_ + (expert_idx * config_.hidden_size + ith * stride) * config_.intermediate_size * ggml_type_size(config_.down_type) / ggml_blck_size(config_.down_type);

2
ktransformers/local_chat.py
View File

@ -32,7 +32,7 @@ custom_models = {
ktransformer_rules_dir = os.path.dirname(os.path.abspath(__file__)) + "/optimize/optimize_rules/" ktransformer_rules_dir = os.path.dirname(os.path.abspath(__file__)) + "/optimize/optimize_rules/"
default_optimize_rules = { default_optimize_rules = {
"DeepseekV2ForCausalLM": ktransformer_rules_dir + "DeepSeek-V2-Chat.yaml", "DeepseekV2ForCausalLM": ktransformer_rules_dir + "DeepSeek-V2-Chat.yaml",
"DeepseekV3ForCausalLM": ktransformer_rules_dir + "DeepSeek-V3-Chat.yaml", "DeepseekV3ForCausalLM": ktransformer_rules_dir + "DeepSeek-V3-Chat-multi-gpu.yaml",
"Qwen2MoeForCausalLM": ktransformer_rules_dir + "Qwen2-57B-A14B-Instruct.yaml", "Qwen2MoeForCausalLM": ktransformer_rules_dir + "Qwen2-57B-A14B-Instruct.yaml",
"LlamaForCausalLM": ktransformer_rules_dir + "Internlm2_5-7b-Chat-1m.yaml", "LlamaForCausalLM": ktransformer_rules_dir + "Internlm2_5-7b-Chat-1m.yaml",
"MixtralForCausalLM": ktransformer_rules_dir + "Mixtral.yaml", "MixtralForCausalLM": ktransformer_rules_dir + "Mixtral.yaml",