vrr/kvcache-ai-ktransformers
2
0
Fork 0
mirror of https://github.com/kvcache-ai/ktransformers.git synced 2025-09-05 20:19:51 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

⚡ update compile option for avx512vpopcntdq

Browse source
This commit is contained in:
liam 2025-03-06 12:18:04 +08:00
parent dd390835ca
commit 8eeb6dd432
5 changed files with 28 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
ktransformers/ktransformers_ext/CMakeLists.txt
View file

@ -175,6 +175,7 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
list(APPEND ARCH_FLAGS -mavx512bw)
list(APPEND ARCH_FLAGS -mavx512dq)
list(APPEND ARCH_FLAGS -mavx512vnni)
list(APPEND ARCH_FLAGS -mavx512vpopcntdq)
endif()
if (LLAMA_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)

14
ktransformers/tests/AIME_2024/eval_api.py
View file

@ -78,13 +78,15 @@ def run_eval_api(
format_tabs: bool = False,
auth_token: str = None,
problem_file: str = None,
append: bool = False
append: bool = False,
skip: int = 0
):
data = load_data(problem_file)
pbar = tqdm.tqdm(total=len(data) * 1)
pbar.update(skip)
for i in range(len(data)):
i = i+skip
data_item = data[i]
question = data_item['Problem']
# Start the timer for this evaluation
@ -97,6 +99,7 @@ def run_eval_api(
score = get_score(completion, answer)
elapsed_time = time.time() - start_time
result = {
"index": i,
"question_id": data_item["ID"],
"answer": answer,
"prediction": completion,
@ -114,9 +117,9 @@ def run_eval_api(
pbar.update(1)
def main(output_path, api_url, model_name, auth_token, format_tabs,problem_file, append):
def main(output_path, api_url, model_name, auth_token, format_tabs,problem_file, append,skip):
os.makedirs(os.path.dirname(output_path), exist_ok=True)
run_eval_api(api_url, model_name, output_path, format_tabs, auth_token, problem_file,append)
run_eval_api(api_url, model_name, output_path, format_tabs, auth_token, problem_file,append,skip)
if __name__ == "__main__":
@ -128,6 +131,7 @@ if __name__ == "__main__":
parser.add_argument("--format_tabs", action="store_true", help="Format Tabs")
parser.add_argument("--problem_file", type=str, default="Maxwell-Jia/AIME_2024", help="Evalset File")
parser.add_argument("--no_append", action="store_false", help="Append to existing file")
parser.add_argument("--skip", type=int, default=0, help="Skip some tasks")
args = parser.parse_args()
# api_url = "https://api.siliconflow.cn/v1/chat/completions"
main(args.out_path, args.api_url, args.model_name, args.auth_token, args.format_tabs, args.problem_file, args.no_append)
main(args.out_path, args.api_url, args.model_name, args.auth_token, args.format_tabs, args.problem_file, args.no_append, args.skip)

21
ktransformers/tests/humaneval/eval_api.py
View file

@ -39,7 +39,8 @@ def run_eval_api(
format_tabs: bool = False,
auth_token: str = None,
problem_file: str = None,
append: bool = False
append: bool = False,
skip: int = 0
):
if(problem_file is None):
problems = read_problems()
@ -47,8 +48,14 @@ def run_eval_api(
problems = read_problems(problem_file)
samples = []
pbar = tqdm.tqdm(total=len(problems) * 1)
pbar.update(skip)
try:
for task_id in problems:
# skip some tasks
if skip > 0:
skip -= 1
continue
if format_tabs:
prompt = problems[task_id]["prompt"].replace(" ", "\t")
else:
@ -67,23 +74,25 @@ def run_eval_api(
if not append:
write_jsonl(out_path, samples,append=append)
except Exception as e:
write_jsonl(out_path, samples,append=append)
if not append:
write_jsonl(out_path, samples,append=append)
print(f"Error: {e}")
def main(output_path, api_url, model_name, auth_token, format_tabs,problem_file, append):
def main(output_path, api_url, model_name, auth_token, format_tabs,problem_file, append,skip):
os.makedirs(os.path.dirname(output_path), exist_ok=True)
run_eval_api(api_url, model_name, output_path, format_tabs, auth_token, problem_file,append)
run_eval_api(api_url, model_name, output_path, format_tabs, auth_token, problem_file,append,skip)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="API Generate Tester")
parser.add_argument("--api_url", type=str, default="https://api.siliconflow.cn/v1/chat/completions", help="API URL")
parser.add_argument("--model_name", type=str, default="Pro/deepseek-ai/DeepSeek-V3", help="Model Name")
parser.add_argument("--out_path", type=str, default="results/api/eval.jsonl", help="Output Path")
parser.add_argument("--out_path", type=str, default="results/api/eval_b.jsonl", help="Output Path")
parser.add_argument("--auth_token", type=str, default=None, help="Auth Token")
parser.add_argument("--format_tabs", action="store_true", help="Format Tabs")
parser.add_argument("--problem_file", type=str, default=None, help="Evalset File")
parser.add_argument("--no_append", action="store_false", help="Append to existing file")
parser.add_argument("--skip", type=int, default=0, help="Skip first n problems")
args = parser.parse_args()
# api_url = "https://api.siliconflow.cn/v1/chat/completions"
main(args.out_path, args.api_url, args.model_name, args.auth_token, args.format_tabs, args.problem_file, args.no_append)
main(args.out_path, args.api_url, args.model_name, args.auth_token, args.format_tabs, args.problem_file, args.no_append,args.skip)

2
ktransformers/tests/humaneval/evaluation.py
View file

@ -8,7 +8,7 @@ def filter_code(completion: str) -> str:
completion = completion.split('if __name__ == "__main__":')[0]
if "# Example usage" in completion:
completion = completion.split("# Example usage")[0]
return completion.split("\n\n")[0]
return completion
def fix_indents(text: str) -> str:

3
third_party/llamafile/iqk_mul_mat.inc vendored
View file

@ -2388,7 +2388,8 @@ struct SimpleBits {
struct EvenSignHelper {
#ifdef HAVE_FANCY_SIMD && __AVX512VPOPCNTDQ__
#if defined HAVE_FANCY_SIMD
// #pragma message("Using AVX512VPOPCNTDQ in even sign helper")
union sbits_t {
__m128i vec;
__mmask32 mask[4];