Spaces:
Sleeping
Sleeping
File size: 8,454 Bytes
e45d058 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 |
import time
import pytest
import torch
from transformers import AutoTokenizer, GPTBigCodeConfig
from transformers.models.gpt_bigcode.modeling_gpt_bigcode import GPTBigCodeForCausalLM
from flash_attn.models.bigcode import bigcode_config_to_gpt2_config, inv_remap_state_dict_hf_bigcode
from flash_attn.models.gpt import GPTLMHeadModel, remap_state_dict_hf_bigcode
from flash_attn.utils.generation import update_graph_cache
from flash_attn.utils.pretrained import state_dict_from_pretrained
@pytest.mark.parametrize("model_name", ["bigcode/starcoderbase-1b", "WizardLM/WizardCoder-1B-V1.0"])
def test_bigcode_state_dict(model_name):
config = bigcode_config_to_gpt2_config(GPTBigCodeConfig.from_pretrained(model_name))
pretrained_state_dict = remap_state_dict_hf_bigcode(
state_dict_from_pretrained(model_name), config
)
model = GPTLMHeadModel(config, device="meta")
state_dict = model.state_dict()
assert state_dict.keys() == pretrained_state_dict.keys()
for k in state_dict.keys():
assert state_dict[k].shape == pretrained_state_dict[k].shape
@pytest.mark.parametrize("model_name", ["bigcode/starcoderbase-1b", "WizardLM/WizardCoder-1B-V1.0"])
def test_bigcode_optimized(model_name):
"""Check that our implementation of BigCode (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
device = "cuda"
config = bigcode_config_to_gpt2_config(GPTBigCodeConfig.from_pretrained(model_name))
config.use_flash_attn = True # FlashAttention-2 supports headdim 256
config.fused_bias_fc = True
config.fused_mlp = True
config.fused_dropout_add_ln = True
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
input_ids = torch.randint(
0, config.vocab_size, (batch_size, max_seqlen), dtype=torch.long, device=device
)
with torch.no_grad():
out = model.transformer(input_ids)
logits = model(input_ids).logits
del model
# Without device_map, the model is loaded on the CPU, which is very slow
model_ref = GPTBigCodeForCausalLM.from_pretrained(model_name, device_map={"": device})
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
model_hf = GPTBigCodeForCausalLM.from_pretrained(
model_name, torch_dtype=dtype, device_map={"": device}
)
model_hf.eval()
out_hf = model_hf.transformer(input_ids).last_hidden_state
logits_hf = model_hf(input_ids).logits
del model_hf
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()}")
assert (out - out_ref).abs().max().item() < 3 * (out_hf - out_ref).abs().max().item()
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()
@pytest.mark.parametrize("model_name", ["bigcode/starcoderbase-1b", "WizardLM/WizardCoder-1B-V1.0"])
def test_bigcode_generation(model_name):
"""Check that our implementation of BigCode (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
device = "cuda"
config = bigcode_config_to_gpt2_config(GPTBigCodeConfig.from_pretrained(model_name))
config.use_flash_attn = True # FlashAttention-2 supports headdim 256
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
input_ids = torch.randint(
0, config.vocab_size, (batch_size, seqlen), dtype=torch.long, device=device
)
model_hf = GPTBigCodeForCausalLM.from_pretrained(
model_name, torch_dtype=dtype, device_map={"": device}
)
model_hf.eval()
print("HF fp16")
torch.cuda.synchronize()
start = time.time()
out_hf = model_hf.generate(
input_ids=input_ids, max_length=max_length, return_dict_in_generate=True, output_scores=True
)
torch.cuda.synchronize()
print(f"Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms")
del model_hf
model_ref = GPTBigCodeForCausalLM.from_pretrained(model_name, device_map={"": device})
model_ref.eval()
with torch.no_grad():
logits_ref = model_ref(out_hf.sequences).logits[:, (seqlen - 1) : -1]
del model_ref
model = GPTLMHeadModel.from_pretrained(model_name, config, device=device, dtype=dtype)
model.eval()
print("Without CUDA graph")
torch.cuda.synchronize()
start = time.time()
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,
enable_timing=True,
teacher_outputs=out_hf.sequences,
)
torch.cuda.synchronize()
print(f"Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms")
# 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)
print("With CUDA graph")
torch.cuda.synchronize()
start = time.time()
out_cg = model.generate(
input_ids=input_ids,
max_length=max_length,
cg=True,
return_dict_in_generate=True,
output_scores=True,
enable_timing=True,
teacher_outputs=out_hf.sequences,
)
torch.cuda.synchronize()
print(f"Prompt processing + decoding time: {(time.time() - start) * 1000:.0f}ms")
with torch.no_grad():
logits_parallel = model(out_hf.sequences).logits[:, (seqlen - 1) : -1]
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
print(f"HF fp16 logits max diff: {hf_error}")
print(f"Logits max diff: {(logits - logits_ref).abs().max().item() }")
assert (logits - logits_ref).abs().max().item() < 2 * hf_error
print(f"Logits CG max diff: {(logits_cg - logits_ref).abs().max().item() }")
assert (logits_cg - logits_ref).abs().max().item() < 2 * hf_error
@pytest.mark.parametrize("model_name", ["bigcode/starcoderbase-1b", "WizardLM/WizardCoder-1B-V1.0"])
def test_inv_remap_state_dict(model_name: str):
"""
Verify that we can convert a HF BigCode model to flash_attn and back.
"""
state_dict = state_dict_from_pretrained(model_name)
config = GPTBigCodeConfig.from_pretrained(model_name)
flash_state_dict = remap_state_dict_hf_bigcode(state_dict, config)
recovered_state_dict = inv_remap_state_dict_hf_bigcode(flash_state_dict, config)
assert set(state_dict.keys()) == set(recovered_state_dict.keys())
for k in state_dict.keys():
assert state_dict[k].shape == recovered_state_dict[k].shape
torch.testing.assert_close(state_dict[k], recovered_state_dict[k], rtol=1e-6, atol=1e-6)
|