ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

[NPU L0] Update 1st token generation (#12314)

Browse source
This commit is contained in:
binbin Deng 2024-11-01 17:02:07 +08:00 committed by GitHub
parent d409d9d0eb
commit f53bb4ea0b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 10 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

23
python/llm/src/ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py
View file

@ -105,18 +105,12 @@ def generate(
key.to(torch.float16).numpy().tofile(os.path.join(temp_dir, f"key_cache_{layer}.bin")) key.to(torch.float16).numpy().tofile(os.path.join(temp_dir, f"key_cache_{layer}.bin"))
val.to(torch.float16).numpy().tofile(os.path.join(temp_dir, f"value_cache_{layer}.bin")) val.to(torch.float16).numpy().tofile(os.path.join(temp_dir, f"value_cache_{layer}.bin"))
token = input_id.to(torch.int32).item()
output_tokens.append(torch.tensor([token]))
if streamer is not None:
streamer.put(torch.tensor([token]))
if "eos_token_id" not in new_generate_kwargs: if "eos_token_id" not in new_generate_kwargs:
eos = 0xffffffff eos = 0xffffffff
else: else:
eos = new_generate_kwargs["eos_token_id"] eos = new_generate_kwargs["eos_token_id"]
time_t1 = time.perf_counter() time_t1 = time.perf_counter()
idx += 1
# start generate_serve by Thread # start generate_serve by Thread
thread = threading.Thread(target=generate_serve, thread = threading.Thread(target=generate_serve,
@ -124,7 +118,7 @@ def generate(
self.head_dim, self.num_layers, self.head_dim, self.num_layers,
self.vocab_size, self.vocab_size,
self.transpose_value_cache, self.transpose_value_cache,
new_tokens - 2)) new_tokens - 1))
thread.start() thread.start()
in_pipe_path = "\\\\.\\pipe\\llminputpipe" in_pipe_path = "\\\\.\\pipe\\llminputpipe"
@ -146,7 +140,7 @@ def generate(
else: else:
break break
time_start = time.perf_counter() time_t2 = time.perf_counter()
bdata = str.encode(str(temp_dir)) bdata = str.encode(str(temp_dir))
invalidInputError(len(bdata) <= 2000, invalidInputError(len(bdata) <= 2000,
@ -162,6 +156,8 @@ def generate(
break break
token = int.from_bytes(data, sys.byteorder) token = int.from_bytes(data, sys.byteorder)
idx += 1 idx += 1
if idx == 1:
time_t3 = time.perf_counter()
if token == eos: if token == eos:
break break
output_tokens.append(torch.tensor([token])) output_tokens.append(torch.tensor([token]))
@ -177,13 +173,14 @@ def generate(
time_end = time.perf_counter() time_end = time.perf_counter()
if do_print: if do_print:
print(f" Start the thread and connect the pipe time: {(time_start - time_t1):.2f} s") print(f" Start the thread and connect the pipe time: {(time_t2 - time_t1):.2f} s")
print(f" Number of input tokens: {input_length}") print(f" Number of input tokens: {input_length}")
print(f" Generated tokens: {idx}") print(f" Generated tokens: {idx}")
print(f" First token generation time: {(time_t1 - time_start_all):.2f} s") print(" First token generation time: "
print(f" Generation average latency: {(time_end - time_start) * 1000 /(idx - 1):.2f} ms, " f"{(time_t3 - time_start_all - (time_t2 - time_t1)):.2f} s")
f"({(idx - 1)/(time_end - time_start):.2f} token/s)") print(f" Generation average latency: {(time_end - time_t3) * 1000 /(idx - 1):.2f} ms, "
print(f" Generation time: {(time_end - time_start_all - (time_start - time_t1)):.2f} s\n") f"({(idx - 1)/(time_end - time_t3):.2f} token/s)")
print(f" Generation time: {(time_end - time_start_all - (time_t2 - time_t1)):.2f} s\n")
return output return output