Migrate logits computation and gather to model_runner (#3233)
This commit is contained in:
@@ -13,6 +13,7 @@ from huggingface_hub import snapshot_download
|
||||
import vllm
|
||||
from vllm.config import LoRAConfig
|
||||
from vllm.model_executor.layers.sampler import Sampler
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.model_loader import get_model
|
||||
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
MergedColumnParallelLinear,
|
||||
@@ -85,7 +86,8 @@ def dummy_model() -> nn.Module:
|
||||
("outact", nn.Sigmoid()),
|
||||
# Special handling for lm_head & sampler
|
||||
("lm_head", ParallelLMHead(512, 10)),
|
||||
("sampler", Sampler(512))
|
||||
("logits_processor", LogitsProcessor(512)),
|
||||
("sampler", Sampler())
|
||||
]))
|
||||
model.config = MagicMock()
|
||||
return model
|
||||
@@ -110,7 +112,8 @@ def dummy_model_gate_up() -> nn.Module:
|
||||
("outact", nn.Sigmoid()),
|
||||
# Special handling for lm_head & sampler
|
||||
("lm_head", ParallelLMHead(512, 10)),
|
||||
("sampler", Sampler(512))
|
||||
("logits_processor", LogitsProcessor(512)),
|
||||
("sampler", Sampler())
|
||||
]))
|
||||
model.config = MagicMock()
|
||||
return model
|
||||
|
||||
+35
-29
@@ -13,14 +13,14 @@ from vllm.lora.layers import (
|
||||
QKVParallelLinearWithLora,
|
||||
VocabParallelEmbeddingWithLoRA,
|
||||
RowParallelLinearWithLoRA,
|
||||
SamplerWithLoRA,
|
||||
LogitsProcessorWithLoRA,
|
||||
LoRAMapping,
|
||||
BaseLayerWithLoRA,
|
||||
)
|
||||
from vllm.lora.models import (LoRALayerWeights, convert_mapping,
|
||||
PackedLoRALayerWeights)
|
||||
from vllm.config import LoRAConfig
|
||||
from vllm.model_executor.layers.sampler import Sampler
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
MergedColumnParallelLinear,
|
||||
RowParallelLinear,
|
||||
@@ -394,7 +394,7 @@ def test_embeddings_with_new_embeddings(dist_init, num_loras, device) -> None:
|
||||
@torch.inference_mode()
|
||||
@pytest.mark.parametrize("num_loras", [1, 2, 4, 8])
|
||||
@pytest.mark.parametrize("device", CUDA_DEVICES)
|
||||
def test_lm_head_sampler(dist_init, num_loras, device) -> None:
|
||||
def test_lm_head_logits_processor(dist_init, num_loras, device) -> None:
|
||||
|
||||
torch.set_default_device(device)
|
||||
max_loras = 8
|
||||
@@ -402,28 +402,29 @@ def test_lm_head_sampler(dist_init, num_loras, device) -> None:
|
||||
max_lora_rank=8,
|
||||
lora_dtype=torch.float16)
|
||||
|
||||
def create_random_sampler_layer():
|
||||
def _pretest():
|
||||
linear = ParallelLMHead(32000 + lora_config.lora_extra_vocab_size,
|
||||
1024, 32000)
|
||||
linear.weight.data = torch.rand_like(linear.weight.data)
|
||||
linear.weight.data[:, 32000:] = 0
|
||||
sampler = Sampler(32000 + lora_config.lora_extra_vocab_size, 32000)
|
||||
lora_sampler = SamplerWithLoRA(sampler, 1024, linear.weight.dtype,
|
||||
linear.weight.device)
|
||||
lora_sampler.create_lora_weights(max_loras, lora_config)
|
||||
logits_processor = LogitsProcessor(
|
||||
32000 + lora_config.lora_extra_vocab_size, 32000)
|
||||
lora_logits_processor = LogitsProcessorWithLoRA(
|
||||
logits_processor, 1024, linear.weight.dtype, linear.weight.device)
|
||||
lora_logits_processor.create_lora_weights(max_loras, lora_config)
|
||||
|
||||
return linear, sampler, lora_sampler
|
||||
return linear, logits_processor, lora_logits_processor
|
||||
|
||||
for i in range(10):
|
||||
set_random_seed(i)
|
||||
|
||||
id_to_index = get_random_id_to_index(num_loras, max_loras)
|
||||
linear, sampler, lora_sampler = create_random_sampler_layer()
|
||||
linear, logits_processor, lora_logits_processor = _pretest()
|
||||
|
||||
# NOTE: all the generated loras share the same embeddings tensor.
|
||||
lora_dict, _ = populate_loras(
|
||||
id_to_index,
|
||||
layer=lora_sampler,
|
||||
layer=lora_logits_processor,
|
||||
layer_weights=linear.weight,
|
||||
generate_embeddings_tensor=1024,
|
||||
)
|
||||
@@ -447,34 +448,37 @@ def test_lm_head_sampler(dist_init, num_loras, device) -> None:
|
||||
32000,
|
||||
lora_config.lora_extra_vocab_size,
|
||||
)
|
||||
lora_sampler.set_mapping(*mapping_info, )
|
||||
lora_logits_processor.set_mapping(*mapping_info, )
|
||||
|
||||
lora_result = lora_sampler._get_logits(hidden_states=torch.cat(inputs),
|
||||
embedding=linear.weight,
|
||||
embedding_bias=None)
|
||||
lora_result = lora_logits_processor._get_logits(
|
||||
hidden_states=torch.cat(inputs),
|
||||
embedding=linear.weight,
|
||||
embedding_bias=None)
|
||||
|
||||
original_weight = linear.weight.clone()
|
||||
|
||||
linear.weight[sampler.org_vocab_size:sampler.org_vocab_size +
|
||||
linear.weight[logits_processor.
|
||||
org_vocab_size:logits_processor.org_vocab_size +
|
||||
embeddings_tensor_len] = embeddings_tensor
|
||||
|
||||
sampler.org_vocab_size = 32000 + lora_config.lora_extra_vocab_size
|
||||
logits_processor.org_vocab_size = (32000 +
|
||||
lora_config.lora_extra_vocab_size)
|
||||
expected_results = []
|
||||
for input_, lora_id in zip(inputs, prompt_mapping):
|
||||
lora = lora_dict[lora_id]
|
||||
result = sampler._get_logits(hidden_states=input_,
|
||||
embedding=linear.weight,
|
||||
embedding_bias=None)
|
||||
result = logits_processor._get_logits(hidden_states=input_,
|
||||
embedding=linear.weight,
|
||||
embedding_bias=None)
|
||||
result[:, 32000 + embeddings_tensor_len:] = float("-inf")
|
||||
result += input_ @ lora.lora_a @ lora.lora_b * lora.scaling
|
||||
expected_results.append(result)
|
||||
expected_result = torch.cat(expected_results)
|
||||
sampler.org_vocab_size = 32000
|
||||
logits_processor.org_vocab_size = 32000
|
||||
|
||||
# Check that resetting the lora weights succeeds
|
||||
|
||||
for slot_idx in range(max_loras):
|
||||
lora_sampler.reset_lora(slot_idx)
|
||||
lora_logits_processor.reset_lora(slot_idx)
|
||||
|
||||
inputs, index_mapping, prompt_mapping = create_random_inputs(
|
||||
active_lora_ids=[0],
|
||||
@@ -488,14 +492,16 @@ def test_lm_head_sampler(dist_init, num_loras, device) -> None:
|
||||
mapping_info = convert_mapping(lora_mapping, id_to_index, max_loras,
|
||||
32000,
|
||||
lora_config.lora_extra_vocab_size)
|
||||
lora_sampler.set_mapping(*mapping_info, )
|
||||
lora_logits_processor.set_mapping(*mapping_info, )
|
||||
|
||||
lora_result = lora_sampler._get_logits(hidden_states=torch.cat(inputs),
|
||||
embedding=original_weight,
|
||||
embedding_bias=None)[:, :32000]
|
||||
expected_result = sampler._get_logits(hidden_states=torch.cat(inputs),
|
||||
embedding=original_weight,
|
||||
embedding_bias=None)
|
||||
lora_result = lora_logits_processor._get_logits(
|
||||
hidden_states=torch.cat(inputs),
|
||||
embedding=original_weight,
|
||||
embedding_bias=None)[:, :32000]
|
||||
expected_result = logits_processor._get_logits(
|
||||
hidden_states=torch.cat(inputs),
|
||||
embedding=original_weight,
|
||||
embedding_bias=None)
|
||||
|
||||
rtol, atol = TOLERANCES[lora_result.dtype]
|
||||
assert torch.allclose(lora_result,
|
||||
|
||||
Reference in New Issue
Block a user