flash-attention/tests/models/test_gptj.py

185 lines
7.1 KiB
Python
Raw Normal View History

2023-04-19 12:43:37 +08:00
# Copyright (c) 2023, Tri Dao.
2023-04-15 07:50:01 +08:00
import time
2023-03-23 07:16:58 +08:00
import pytest
2023-08-19 11:59:35 +08:00
import torch
2023-03-23 07:16:58 +08:00
from flash_attn.models.gpt import GPTLMHeadModel
2023-08-19 11:59:35 +08:00
from flash_attn.models.gptj import gptj_config_to_gpt2_config, remap_state_dict_hf_gptj
2023-04-15 07:50:01 +08:00
from flash_attn.utils.generation import update_graph_cache
2023-08-19 11:59:35 +08:00
from flash_attn.utils.pretrained import state_dict_from_pretrained
from transformers import AutoTokenizer, GPTJConfig
from transformers.models.gptj.modeling_gptj import GPTJForCausalLM
2023-03-23 07:16:58 +08:00
2023-08-19 11:59:35 +08:00
@pytest.mark.parametrize("model_name", ["EleutherAI/gpt-j-6B"])
2023-03-23 07:16:58 +08:00
def test_gptj_state_dict(model_name):
config = gptj_config_to_gpt2_config(GPTJConfig.from_pretrained(model_name))
pretrained_state_dict = remap_state_dict_hf_gptj(state_dict_from_pretrained(model_name), config)
2023-08-19 11:59:35 +08:00
model = GPTLMHeadModel(config, device="meta") # Without device='meta' init is very slow
2023-03-23 07:16:58 +08:00
state_dict = model.state_dict()
assert state_dict.keys() == pretrained_state_dict.keys()
for k in state_dict.keys():
2023-03-23 07:16:58 +08:00
assert state_dict[k].shape == pretrained_state_dict[k].shape
@pytest.mark.parametrize("model_name", ["EleutherAI/gpt-j-6B", "togethercomputer/GPT-JT-6B-v1"])
2023-03-23 07:16:58 +08:00
def test_gptj_optimized(model_name):
"""Check that our implementation of GPT-J (with all optimizations enabled) matches the
HF implementation: the output of our forward pass in fp16 should be around the same as the HF
forward pass in fp16, when compared to the HF forward pass in fp32.
"""
dtype = torch.float16
2023-08-19 11:59:35 +08:00
device = "cuda"
2023-03-23 07:16:58 +08:00
config = gptj_config_to_gpt2_config(GPTJConfig.from_pretrained(model_name))
2023-07-22 08:29:10 +08:00
config.use_flash_attn = True # FlashAttention-2 supports headdim 256
2023-03-23 07:16:58 +08:00
config.fused_bias_fc = True
config.fused_mlp = True
config.fused_dropout_add_ln = True
2023-03-23 07:16:58 +08:00
config.residual_in_fp32 = True
model = GPTLMHeadModel.from_pretrained(model_name, config, device=device, dtype=dtype)
model.eval()
torch.manual_seed(0)
batch_size = 2
max_seqlen = 256
2023-08-19 11:59:35 +08:00
input_ids = torch.randint(
0, config.vocab_size, (batch_size, max_seqlen), dtype=torch.long, device=device
)
2023-03-23 07:16:58 +08:00
with torch.no_grad():
out = model.transformer(input_ids)
logits = model(input_ids).logits
del model
2023-03-29 16:21:25 +08:00
# Without device_map, the model is loaded on the CPU, which is very slow
model_ref = GPTJForCausalLM.from_pretrained(model_name, device_map={"": device})
2023-03-23 07:16:58 +08:00
model_ref.eval()
with torch.no_grad():
out_ref = model_ref.transformer(input_ids).last_hidden_state
logits_ref = model_ref(input_ids).logits
del model_ref
2023-08-19 11:59:35 +08:00
model_hf = GPTJForCausalLM.from_pretrained(
model_name, torch_dtype=dtype, device_map={"": device}
)
2023-03-23 07:16:58 +08:00
model_hf.eval()
out_hf = model_hf.transformer(input_ids).last_hidden_state
logits_hf = model_hf(input_ids).logits
del model_hf
2023-08-19 11:59:35 +08:00
print(f"Output max diff: {(out - out_ref).abs().max().item()}")
print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
print(f"HF fp16 max diff: {(out_hf - out_ref).abs().max().item()}")
print(f"HF fp16 mean diff: {(out_hf - out_ref).abs().mean().item()}")
2023-03-23 07:16:58 +08:00
assert (out - out_ref).abs().max().item() < 3 * (out_hf - out_ref).abs().max().item()
2023-08-19 11:59:35 +08:00
print(f"Logits max diff: {(logits - logits_ref).abs().max().item()}")
print(f"Logits mean diff: {(logits - logits_ref).abs().mean().item()}")
print(f"HF fp16 max diff: {(logits_hf - logits_ref).abs().max().item()}")
print(f"HF fp16 mean diff: {(logits_hf - logits_ref).abs().mean().item()}")
assert (logits - logits_ref).abs().max().item() < 3 * (
logits_hf - logits_ref
).abs().max().item()
2023-04-15 07:50:01 +08:00
2023-08-19 11:59:35 +08:00
@pytest.mark.parametrize("model_name", ["EleutherAI/gpt-j-6B"])
def test_gptj_generation(model_name):
2023-04-15 07:50:01 +08:00
"""Check that our implementation of GPT-J (with all optimizations enabled) matches the
HF implementation: the output of our forward pass in fp16 should be around the same as the HF
forward pass in fp16, when compared to the HF forward pass in fp32.
"""
dtype = torch.float16
2023-08-19 11:59:35 +08:00
device = "cuda"
2023-04-15 07:50:01 +08:00
config = gptj_config_to_gpt2_config(GPTJConfig.from_pretrained(model_name))
2023-07-22 08:29:10 +08:00
config.use_flash_attn = True # FlashAttention-2 supports headdim 256
2023-04-15 07:50:01 +08:00
config.fused_bias_fc = True
config.fused_mlp = True
config.fused_dropout_add_ln = True
# Only prenorm supports residual_in_fp32
config.residual_in_fp32 = True
tokenizer = AutoTokenizer.from_pretrained(model_name)
eos_token_id = tokenizer.eos_token_id
torch.manual_seed(0)
batch_size = 1
seqlen = 100
max_length = 150
2023-08-19 11:59:35 +08:00
input_ids = torch.randint(
0, config.vocab_size, (batch_size, seqlen), dtype=torch.long, device=device
)
2023-04-15 07:50:01 +08:00
2023-08-19 11:59:35 +08:00
model_hf = GPTJForCausalLM.from_pretrained(
model_name, torch_dtype=dtype, device_map={"": device}
)
2023-04-15 07:50:01 +08:00
model_hf.eval()
print("HF fp16")
torch.cuda.synchronize()
start = time.time()
2023-08-19 11:59:35 +08:00
out_hf = model_hf.generate(
input_ids=input_ids, max_length=max_length, return_dict_in_generate=True, output_scores=True
)
2023-04-15 07:50:01 +08:00
torch.cuda.synchronize()
2023-08-19 11:59:35 +08:00
print(f"Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms")
2023-04-15 07:50:01 +08:00
del model_hf
model_ref = GPTJForCausalLM.from_pretrained(model_name, device_map={"": device})
model_ref.eval()
with torch.no_grad():
2023-08-19 11:59:35 +08:00
logits_ref = model_ref(out_hf.sequences).logits[:, (seqlen - 1) : -1]
2023-04-15 07:50:01 +08:00
del model_ref
model = GPTLMHeadModel.from_pretrained(model_name, config, device=device, dtype=dtype)
model.eval()
2023-08-19 11:59:35 +08:00
print("Without CUDA graph")
2023-04-15 07:50:01 +08:00
torch.cuda.synchronize()
start = time.time()
2023-08-19 11:59:35 +08:00
out = model.generate(
input_ids=input_ids,
max_length=max_length,
eos_token_id=eos_token_id,
return_dict_in_generate=True,
output_scores=True,
2023-09-05 08:01:38 +08:00
enable_timing=True,
2023-08-19 11:59:35 +08:00
teacher_outputs=out_hf.sequences,
)
2023-04-15 07:50:01 +08:00
torch.cuda.synchronize()
2023-08-19 11:59:35 +08:00
print(f"Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms")
2023-04-15 07:50:01 +08:00
# Capture graph outside the timing loop
batch_size, seqlen_og = input_ids.shape
model._decoding_cache = update_graph_cache(model, None, batch_size, seqlen_og, max_length)
2023-08-19 11:59:35 +08:00
print("With CUDA graph")
2023-04-15 07:50:01 +08:00
torch.cuda.synchronize()
start = time.time()
2023-08-19 11:59:35 +08:00
out_cg = model.generate(
input_ids=input_ids,
max_length=max_length,
cg=True,
return_dict_in_generate=True,
output_scores=True,
2023-09-05 08:01:38 +08:00
enable_timing=True,
2023-08-19 11:59:35 +08:00
teacher_outputs=out_hf.sequences,
)
2023-04-15 07:50:01 +08:00
torch.cuda.synchronize()
2023-08-19 11:59:35 +08:00
print(f"Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms")
2023-04-15 07:50:01 +08:00
with torch.no_grad():
2023-08-19 11:59:35 +08:00
logits_parallel = model(out_hf.sequences).logits[:, (seqlen - 1) : -1]
2023-04-15 07:50:01 +08:00
logits_hf = torch.stack(out_hf.scores, dim=1)
logits = torch.stack(out.scores, dim=1)
logits_cg = torch.stack(out_cg.scores, dim=1)
del model
hf_error = (logits_hf - logits_ref).abs().max().item()
assert (logits_parallel - logits_ref).abs().max().item() < 2 * hf_error
2023-08-19 11:59:35 +08:00
print(f"HF fp16 logits max diff: {hf_error}")
print(f"Logits max diff: {(logits - logits_ref).abs().max().item() }")
2023-04-15 07:50:01 +08:00
assert (logits - logits_ref).abs().max().item() < 2 * hf_error
2023-08-19 11:59:35 +08:00
print(f"Logits CG max diff: {(logits_cg - logits_ref).abs().max().item() }")
2023-04-15 07:50:01 +08:00
assert torch.equal(logits_cg, logits)