vrr/kvcache-ai-ktransformers
2
0
Fork 0
mirror of https://github.com/kvcache-ai/ktransformers.git synced 2025-09-15 01:29:42 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

⚡ add flush print in local_chat output and change default optimize yaml of deepseekv3 to single gpu

Browse source
This commit is contained in:
liam 2025-02-08 13:15:52 +08:00
parent 0262f954c7
commit c18ecd7b7f
2 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

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-multi-gpu.yaml", "DeepseekV3ForCausalLM": ktransformer_rules_dir + "DeepSeek-V3-Chat.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",

4
ktransformers/server/backend/interfaces/transformers.py
View file

@ -329,14 +329,14 @@ class TransformersInterface(BackendInterfaceBase):
self.profiler.create_and_start_timer("prefill") self.profiler.create_and_start_timer("prefill")
for t in self.prefill(input_ids, self.check_is_new(thread_id)): for t in self.prefill(input_ids, self.check_is_new(thread_id)):
if t is not None: if t is not None:
print(t, end="") print(t, end="",flush=True)
yield t yield t
self.profiler.pause_timer("prefill") self.profiler.pause_timer("prefill")
self.profiler.create_and_start_timer("decode") self.profiler.create_and_start_timer("decode")
for t in self.generate(): for t in self.generate():
if t is not None: if t is not None:
print(t, end="") print(t, end="",flush=True)
yield t yield t
print("") print("")
self.profiler.pause_timer("decode") self.profiler.pause_timer("decode")