[Kernel][RFC] Refactor the punica kernel based on Triton (#5036)

This commit is contained in:
Jee Jee Li
2024-07-31 17:12:24 -07:00
committed by GitHub
parent 7eb0cb4a14
commit 7ecee34321
47 changed files with 3177 additions and 4366 deletions
+30 -18
View File
@@ -1,14 +1,17 @@
import gc
from unittest.mock import patch
import pytest
import torch
import triton
import triton.language as tl
from vllm.model_executor.layers.ops.sample import (_uniform_to_exponential,
from vllm.model_executor.layers.ops.sample import (_sample_triton,
_uniform_to_exponential,
sample)
from vllm.model_executor.sampling_metadata import SamplingTensors
from vllm.model_executor.utils import set_random_seed
from vllm.triton_utils.libentry import LibEntry
from vllm.triton_utils.sample import (MAX_TRITON_N_COLS,
get_num_triton_sampler_splits)
@@ -76,15 +79,20 @@ def test_sample_decoding_only(random_sampling, max_best_of,
seeds = torch.randint(1,
torch.iinfo(torch.long).max, (n_splits, bs),
device="cuda").mul_(random_sampling_mask)
sampled_tokens, sampled_logprobs, sampled_modified_probs = sample(
probs=probs,
logprobs=logprobs,
sample_indices=sample_indices,
seeds=seeds,
max_best_of=max_best_of,
modify_greedy_probs=modify_greedy_probs,
save_logprobs=save_logprobs,
_save_modified_probs=True)
#The current _sample_triton does not utilize the
# libentry decoration. The purpose of adding this patch is to test
# the correctness of libentry.
with patch("vllm.model_executor.layers.ops.sample._sample_triton",
LibEntry(_sample_triton)):
sampled_tokens, sampled_logprobs, sampled_modified_probs = sample(
probs=probs,
logprobs=logprobs,
sample_indices=sample_indices,
seeds=seeds,
max_best_of=max_best_of,
modify_greedy_probs=modify_greedy_probs,
save_logprobs=save_logprobs,
_save_modified_probs=True)
assert sampled_tokens.shape == (bs, max_best_of)
for i in range(bs):
assert torch.all(sampled_tokens[i] == i * (vocab_size // bs))
@@ -130,6 +138,7 @@ def test_sample_decoding_only(random_sampling, max_best_of,
[SINGLE_SPLIT_VOCAB_SIZE, MULTI_SPLIT_VOCAB_SIZE])
def test_sample_prompt_logprobs(random_sampling, max_best_of,
modify_greedy_probs, seed, vocab_size):
set_random_seed(seed)
prompt_sizes = [16, 32, 64, 128] * 2
samples = 8
@@ -157,14 +166,17 @@ def test_sample_prompt_logprobs(random_sampling, max_best_of,
seeds = torch.randint(1,
torch.iinfo(torch.long).max, (n_splits, samples),
device="cuda").mul_(random_sampling_mask)
sampled_tokens, sampled_logprobs, _ = sample(
probs=probs,
logprobs=logprobs,
sample_indices=sample_indices,
seeds=seeds,
max_best_of=max_best_of,
modify_greedy_probs=modify_greedy_probs,
save_logprobs=True)
#ditto
with patch("vllm.model_executor.layers.ops.sample._sample_triton",
LibEntry(_sample_triton)):
sampled_tokens, sampled_logprobs, _ = sample(
probs=probs,
logprobs=logprobs,
sample_indices=sample_indices,
seeds=seeds,
max_best_of=max_best_of,
modify_greedy_probs=modify_greedy_probs,
save_logprobs=True)
assert sampled_tokens.shape == (samples, max_best_of)
assert sampled_logprobs.shape == (samples, max_best_of)
for i, t in enumerate(sample_indices):
+1 -1
View File
@@ -37,7 +37,7 @@ def test_gemma_lora(gemma_lora_files):
expected_lora_output = [
"more important than knowledge.\nAuthor: Albert Einstein\n",
"everyone else is already taken.\nAuthor: Oscar Wilde\n",
"so little time\nAuthor: Frank Zappa\n",
"so little time.\nAuthor: Frank Zappa\n",
]
output1 = do_sample(llm, gemma_lora_files, lora_id=1)
+81 -59
View File
@@ -26,7 +26,8 @@ from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,
VocabParallelEmbeddingWithLoRA)
# yapf: enable
from vllm.lora.models import (LongContextLoRAContext, LoRALayerWeights,
PackedLoRALayerWeights, convert_mapping)
PackedLoRALayerWeights)
from vllm.lora.punica import PunicaWrapper
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
MergedColumnParallelLinear,
QKVParallelLinear,
@@ -47,6 +48,9 @@ TOLERANCES = {
CUDA_DEVICES = [
f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
]
# We will launch different triton kernels between the prefill and decode
# stages, so we need to verify this. prefill stage(True) or decode stage(False)
STAGES = [True, False]
def get_random_id_to_index(num_loras: int,
@@ -182,10 +186,12 @@ def create_random_inputs(
@pytest.mark.parametrize("num_loras", [1, 2, 4, 8])
@pytest.mark.parametrize("device", CUDA_DEVICES)
@pytest.mark.parametrize("vocab_size", [512, 32000, 64000, 128000])
def test_embeddings(dist_init, num_loras, device, vocab_size) -> None:
@pytest.mark.parametrize("stage", STAGES)
def test_embeddings(dist_init, num_loras, device, vocab_size, stage) -> None:
torch.set_default_device(device)
max_loras = 8
punica_wrapper = PunicaWrapper(8192, 256, device)
lora_config = LoRAConfig(max_loras=max_loras,
max_lora_rank=8,
lora_dtype=torch.float16)
@@ -204,7 +210,7 @@ def test_embeddings(dist_init, num_loras, device, vocab_size) -> None:
id_to_index = get_random_id_to_index(num_loras, max_loras)
embedding, lora_embedding = create_random_embedding_layer()
lora_embedding.set_mapping(punica_wrapper)
lora_dict, _ = populate_loras(
id_to_index,
layer=lora_embedding,
@@ -217,12 +223,12 @@ def test_embeddings(dist_init, num_loras, device, vocab_size) -> None:
input_size=(200, ),
input_range=(1, vocab_size),
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
mapping_info = convert_mapping(lora_mapping, id_to_index, max_loras,
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
punica_wrapper.update_metadata(lora_mapping, id_to_index, max_loras,
vocab_size,
lora_config.lora_extra_vocab_size)
lora_embedding.set_mapping(*mapping_info)
lora_result = lora_embedding(torch.cat(inputs))
@@ -255,12 +261,12 @@ def test_embeddings(dist_init, num_loras, device, vocab_size) -> None:
input_size=(200, ),
input_range=(1, vocab_size),
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
mapping_info = convert_mapping(lora_mapping, id_to_index, max_loras,
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
punica_wrapper.update_metadata(lora_mapping, id_to_index, max_loras,
vocab_size,
lora_config.lora_extra_vocab_size)
lora_embedding.set_mapping(*mapping_info, )
lora_result = lora_embedding(torch.cat(inputs))
expected_result = embedding(torch.cat(inputs))
@@ -278,11 +284,13 @@ def test_embeddings(dist_init, num_loras, device, vocab_size) -> None:
@pytest.mark.parametrize("num_loras", [1, 2, 4, 8])
@pytest.mark.parametrize("device", CUDA_DEVICES)
@pytest.mark.parametrize("vocab_size", [512, 32000, 64000, 128000])
@pytest.mark.parametrize("stage", STAGES)
def test_embeddings_with_new_embeddings(dist_init, num_loras, device,
vocab_size) -> None:
vocab_size, stage) -> None:
torch.set_default_device(device)
max_loras = 8
punica_wrapper = PunicaWrapper(8192, 256, device)
lora_config = LoRAConfig(max_loras=max_loras,
max_lora_rank=8,
lora_dtype=torch.float16)
@@ -318,6 +326,7 @@ def test_embeddings_with_new_embeddings(dist_init, num_loras, device,
generate_embeddings_tensor=256,
)
lora_embedding.set_mapping(punica_wrapper)
# All embeddings tensors have the same shape.
embeddings_tensors = [
lora_dict[id].embeddings_tensor for id in sorted(lora_dict.keys())
@@ -334,8 +343,12 @@ def test_embeddings_with_new_embeddings(dist_init, num_loras, device,
input_size=(200, ),
input_range=(1, vocab_size),
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
punica_wrapper.update_metadata(lora_mapping, id_to_index, max_loras,
vocab_size,
lora_config.lora_extra_vocab_size)
original_inputs = deepcopy(inputs)
# Force some of the inputs to be in the extended embeddings range
@@ -349,11 +362,6 @@ def test_embeddings_with_new_embeddings(dist_init, num_loras, device,
(embedding_id + 1) * embeddings_tensor_len - 1)
original_input_[-2] = vocab_size + embeddings_tensor_len - 1
mapping_info = convert_mapping(lora_mapping, id_to_index, max_loras,
vocab_size,
lora_config.lora_extra_vocab_size)
lora_embedding.set_mapping(*mapping_info, )
expanded_embedding.weight[vocab_size:vocab_size +
(embeddings_tensor_len *
max_loras)] = torch.cat(embeddings_tensors)
@@ -390,15 +398,13 @@ def test_embeddings_with_new_embeddings(dist_init, num_loras, device,
input_size=(200, ),
input_range=(1, vocab_size),
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
original_inputs = deepcopy(inputs)
mapping_info = convert_mapping(lora_mapping, id_to_index, max_loras,
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
punica_wrapper.update_metadata(lora_mapping, id_to_index, max_loras,
vocab_size,
lora_config.lora_extra_vocab_size)
lora_embedding.set_mapping(*mapping_info, )
lora_result = lora_embedding(torch.cat(original_inputs))
expected_result = expanded_embedding(torch.cat(inputs))
@@ -413,11 +419,13 @@ def test_embeddings_with_new_embeddings(dist_init, num_loras, device,
@pytest.mark.parametrize("num_loras", [1, 2, 4, 8])
@pytest.mark.parametrize("device", CUDA_DEVICES)
@pytest.mark.parametrize("vocab_size", [512, 32000, 64000, 128000])
def test_lm_head_logits_processor(dist_init, num_loras, device,
vocab_size) -> None:
@pytest.mark.parametrize("stage", STAGES)
def test_lm_head_logits_processor(dist_init, num_loras, device, vocab_size,
stage) -> None:
torch.set_default_device(device)
max_loras = 8
punica_wrapper = PunicaWrapper(8192, 256, device)
lora_config = LoRAConfig(max_loras=max_loras,
max_lora_rank=8,
lora_dtype=torch.float16)
@@ -443,7 +451,7 @@ def test_lm_head_logits_processor(dist_init, num_loras, device,
id_to_index = get_random_id_to_index(num_loras, max_loras)
linear, logits_processor, lora_logits_processor = _pretest()
lora_logits_processor.set_mapping(punica_wrapper)
# NOTE: all the generated loras share the same embeddings tensor.
lora_dict, _ = populate_loras(
id_to_index,
@@ -461,17 +469,17 @@ def test_lm_head_logits_processor(dist_init, num_loras, device,
input_range=(0, 1),
input_type=torch.float16,
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
input_ = torch.rand(20, 1024)
mapping_info = convert_mapping(
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
punica_wrapper.update_metadata(
lora_mapping,
id_to_index,
max_loras,
vocab_size,
lora_config.lora_extra_vocab_size,
)
lora_logits_processor.set_mapping(*mapping_info, )
input_ = torch.rand(20, 1024)
lora_result = lora_logits_processor._get_logits(
hidden_states=torch.cat(inputs),
@@ -510,12 +518,16 @@ def test_lm_head_logits_processor(dist_init, num_loras, device,
input_range=(0, 1),
input_type=torch.float16,
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
mapping_info = convert_mapping(lora_mapping, id_to_index, max_loras,
vocab_size,
lora_config.lora_extra_vocab_size)
lora_logits_processor.set_mapping(*mapping_info, )
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
punica_wrapper.update_metadata(
lora_mapping,
id_to_index,
max_loras,
vocab_size,
lora_config.lora_extra_vocab_size,
)
lora_result = lora_logits_processor._get_logits(
hidden_states=torch.cat(inputs),
@@ -538,10 +550,12 @@ def test_lm_head_logits_processor(dist_init, num_loras, device,
@pytest.mark.parametrize("orientation", ["row", "column"])
@pytest.mark.parametrize("fully_shard", [True, False])
@pytest.mark.parametrize("device", CUDA_DEVICES)
@pytest.mark.parametrize("stage", STAGES)
def test_linear_parallel(dist_init, num_loras, orientation, fully_shard,
device) -> None:
device, stage) -> None:
torch.set_default_device(device)
punica_wrapper = PunicaWrapper(8192, 256, device)
max_loras = 8
lora_config = LoRAConfig(max_loras=max_loras,
max_lora_rank=8,
@@ -575,7 +589,7 @@ def test_linear_parallel(dist_init, num_loras, orientation, fully_shard,
id_to_index = get_random_id_to_index(num_loras, max_loras)
linear, lora_linear = create_random_linear_parallel_layer()
lora_linear.set_mapping(punica_wrapper)
lora_dict, _ = populate_loras(
id_to_index,
layer=lora_linear,
@@ -589,16 +603,16 @@ def test_linear_parallel(dist_init, num_loras, orientation, fully_shard,
input_range=(0, 1),
input_type=torch.float16,
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
mapping_info = convert_mapping(
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
punica_wrapper.update_metadata(
lora_mapping,
id_to_index,
max_loras,
512,
lora_config.lora_extra_vocab_size,
)
lora_linear.set_mapping(*mapping_info, )
lora_result = lora_linear(torch.cat(inputs))[0]
@@ -628,11 +642,12 @@ def test_linear_parallel(dist_init, num_loras, orientation, fully_shard,
input_range=(0, 1),
input_type=torch.float16,
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
mapping_info = convert_mapping(lora_mapping, id_to_index, max_loras,
punica_wrapper.update_metadata(lora_mapping, id_to_index, max_loras,
512, lora_config.lora_extra_vocab_size)
lora_linear.set_mapping(*mapping_info, )
lora_result = lora_linear(torch.cat(inputs))[0]
expected_result = linear(torch.cat(inputs))[0]
@@ -649,10 +664,12 @@ def test_linear_parallel(dist_init, num_loras, orientation, fully_shard,
@pytest.mark.parametrize("repeats", [1, 2, 3])
@pytest.mark.parametrize("fully_shard", [True, False])
@pytest.mark.parametrize("device", CUDA_DEVICES)
@pytest.mark.parametrize("stage", STAGES)
def test_column_parallel_packed(dist_init, num_loras, repeats, fully_shard,
device) -> None:
device, stage) -> None:
torch.set_default_device(device)
punica_wrapper = PunicaWrapper(8192, 256, device)
max_loras = 8
lora_config = LoRAConfig(max_loras=max_loras,
max_lora_rank=8,
@@ -707,7 +724,7 @@ def test_column_parallel_packed(dist_init, num_loras, repeats, fully_shard,
id_to_index = get_random_id_to_index(num_loras, max_loras)
linear, lora_linear = create_column_parallel_packed_layer()
lora_linear.set_mapping(punica_wrapper)
lora_dict, sublora_dict = populate_loras(
id_to_index,
layer=lora_linear,
@@ -722,16 +739,17 @@ def test_column_parallel_packed(dist_init, num_loras, repeats, fully_shard,
input_range=(0, 1),
input_type=torch.float16,
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
mapping_info = convert_mapping(
punica_wrapper.update_metadata(
lora_mapping,
id_to_index,
max_loras,
512,
lora_config.lora_extra_vocab_size,
)
lora_linear.set_mapping(*mapping_info)
lora_result = lora_linear(torch.cat(inputs))[0]
@@ -762,16 +780,18 @@ def test_column_parallel_packed(dist_init, num_loras, repeats, fully_shard,
input_range=(0, 1),
input_type=torch.float16,
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
lora_mapping = LoRAMapping(index_mapping,
prompt_mapping,
is_prefill=stage)
mapping_info = convert_mapping(
punica_wrapper.update_metadata(
lora_mapping,
id_to_index,
max_loras,
512,
lora_config.lora_extra_vocab_size,
)
lora_linear.set_mapping(*mapping_info)
# lora_linear.set_mapping(*mapping_info)
lora_result = lora_linear(torch.cat(inputs))[0]
expected_result = linear(torch.cat(inputs))[0]
@@ -803,7 +823,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
torch.set_default_device(device)
punica_wrapper = PunicaWrapper(8192, 256, device)
max_loras = 8
lora_config = LoRAConfig(max_loras=max_loras,
max_lora_rank=8,
@@ -825,6 +845,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
is_neox_style,
)
lora_rope = LinearScalingRotaryEmbeddingWithLora(rope)
lora_rope.set_mapping(punica_wrapper)
lora_rope.create_lora_weights(max_loras, lora_config)
linear_rope = get_rope(head_size, rotary_dim, max_position, base,
is_neox_style, {
@@ -840,6 +861,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
input_range=(0, lora_config.lora_extra_vocab_size),
input_type=torch.float16,
)
lora_mapping = LoRAMapping(index_mapping, prompt_mapping)
long_lora_context = LongContextLoRAContext(list(scaling_factors),
rotary_dim)
@@ -854,7 +876,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
for i in range(len(scaling_factors)):
long_lora_context.offsets_by_lora_id[i] = scaling_factor_to_offset.get(
scaling_factors[i], 0)
mapping_info = convert_mapping(
punica_wrapper.update_metadata(
lora_mapping,
id_to_index,
max_loras,
@@ -862,7 +884,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
lora_config.lora_extra_vocab_size,
long_lora_context=long_lora_context,
)
lora_rope.set_mapping(*mapping_info)
# lora_rope.set_mapping(*mapping_info)
positions = torch.randint(0, max_position, (batch_size, seq_len))
query = torch.randn(batch_size,
-224
View File
@@ -1,224 +0,0 @@
import pytest
import torch
from vllm.lora.layers import _apply_lora, _apply_lora_packed_nslice
from .utils import DummyLoRAManager
TENSOR_SIZES = [128, 1024, 2048, 4096, 8192, 11008, 11008 // 2, 11008 // 4]
QKV_TENSOR_SIZES = [
(8192, 1024, 1024),
(8192 // 8, 1024 // 8, 1024 // 8),
(4096, 4096, 4096),
(4096 // 2, 4096 // 2, 4096 // 2),
]
BATCH_SIZES = [8, 32, 256]
RANKS = [8]
DTYPES = [torch.float16]
TOLERANCES = {
torch.float16: (5e-3, 5e-3),
torch.bfloat16: (3e-2, 2e-2),
}
@pytest.mark.parametrize("m", TENSOR_SIZES)
@pytest.mark.parametrize("n", TENSOR_SIZES)
@pytest.mark.parametrize("k", BATCH_SIZES)
@pytest.mark.parametrize("rank", RANKS)
@pytest.mark.parametrize("dtype", DTYPES)
def test_apply_lora(m, n, k, rank, dtype) -> None:
manager = DummyLoRAManager()
module_name = "module"
weight = torch.rand([m, n], device="cuda", dtype=dtype)
manager.init_random_lora(module_name, weight, rank=rank)
lora = manager.get_module_lora(module_name)
input = torch.rand(k, n, device="cuda", dtype=dtype)
expected = input @ lora.lora_a @ lora.lora_b * lora.scaling
lora_a_stack = torch.zeros(8,
1,
lora.lora_a.shape[1],
lora.lora_a.shape[0],
device="cuda",
dtype=dtype)
lora_b_stack = torch.zeros(8,
1,
lora.lora_b.shape[1],
lora.lora_b.shape[0],
device="cuda",
dtype=dtype)
for i in range(lora_a_stack.shape[0]):
lora_a_stack[i][0] = lora.lora_a.T
lora_b_stack[i][0] = (lora.lora_b * lora.scaling).T
output = torch.zeros(k, m, device="cuda", dtype=dtype)
_apply_lora(
input, lora_a_stack, lora_b_stack,
torch.randint(0, lora_a_stack.shape[0], (len(input), ), device="cuda"),
output)
rtol, atol = TOLERANCES[dtype]
assert torch.allclose(expected, output, rtol=rtol, atol=atol)
output[:] = 0
_apply_lora(input, lora_a_stack, lora_b_stack,
torch.full((len(input), ), -1, device="cuda"), output)
assert torch.allclose(torch.zeros_like(output), output)
manager.reset_lora()
@pytest.mark.parametrize("m", TENSOR_SIZES)
@pytest.mark.parametrize("n", TENSOR_SIZES)
@pytest.mark.parametrize("k", BATCH_SIZES)
@pytest.mark.parametrize("rank", RANKS)
@pytest.mark.parametrize("dtype", DTYPES)
def test_apply_lora_packed_2slice(m, n, k, rank, dtype) -> None:
if m % 2 != 0:
pytest.skip("m must be divisible by 2")
if m // 2 not in TENSOR_SIZES:
pytest.skip("m//2 must be in TENSOR_SIZES")
manager = DummyLoRAManager()
module_name = "module"
weight = torch.rand([m // 2, n], device="cuda", dtype=dtype)
manager.init_random_lora(module_name + "1", weight, rank=rank)
lora_1 = manager.get_module_lora(module_name + "1")
manager.init_random_lora(module_name + "2", weight, rank=rank)
lora_2 = manager.get_module_lora(module_name + "2")
input = torch.rand(k, n, device="cuda", dtype=dtype)
expected = torch.cat([
input @ lora_1.lora_a @ lora_1.lora_b * lora_1.scaling,
input @ lora_2.lora_a @ lora_2.lora_b * lora_2.scaling
],
dim=1)
lora_a_stacks = [
torch.zeros(8,
1,
lora_1.lora_a.shape[1],
lora_1.lora_a.shape[0],
device="cuda",
dtype=dtype) for i in range(2)
]
lora_b_stacks = [
torch.zeros(8,
1,
lora_1.lora_b.shape[1],
lora_1.lora_b.shape[0],
device="cuda",
dtype=dtype) for i in range(2)
]
for i in range(lora_a_stacks[0].shape[0]):
lora_a_stacks[0][i][0] = lora_1.lora_a.T
lora_b_stacks[0][i][0] = (lora_1.lora_b * lora_1.scaling).T
lora_a_stacks[1][i][0] = lora_2.lora_a.T
lora_b_stacks[1][i][0] = (lora_2.lora_b * lora_2.scaling).T
output = torch.zeros(k, m, device="cuda", dtype=dtype)
_apply_lora_packed_nslice(
input, lora_a_stacks, lora_b_stacks,
torch.randint(0,
lora_a_stacks[0].shape[0], (len(input), ),
device="cuda"), output, (m // 2, m // 2))
rtol, atol = TOLERANCES[dtype]
assert torch.allclose(expected, output, rtol=rtol, atol=atol)
output[:] = 0
_apply_lora_packed_nslice(input, lora_a_stacks, lora_b_stacks,
torch.full((len(input), ), -1, device="cuda"),
output, (m // 2, m // 2))
assert torch.allclose(torch.zeros_like(output), output)
manager.reset_lora()
@pytest.mark.parametrize("qkv", QKV_TENSOR_SIZES)
@pytest.mark.parametrize("n", TENSOR_SIZES)
@pytest.mark.parametrize("k", BATCH_SIZES)
@pytest.mark.parametrize("rank", RANKS)
@pytest.mark.parametrize("dtype", DTYPES)
def test_apply_lora_packed_3slice(qkv, n, k, rank, dtype) -> None:
manager = DummyLoRAManager()
module_name = "module"
weight_q = torch.empty(qkv[0], n, device="cuda", dtype=dtype)
weight_kv = torch.empty(qkv[1], n, device="cuda", dtype=dtype)
manager.init_random_lora(module_name + "q", weight_q, rank=rank)
lora_q = manager.get_module_lora(module_name + "q")
manager.init_random_lora(module_name + "k", weight_kv, rank=rank)
lora_k = manager.get_module_lora(module_name + "k")
manager.init_random_lora(module_name + "v", weight_kv, rank=rank)
lora_v = manager.get_module_lora(module_name + "v")
input = torch.rand(k, n, device="cuda", dtype=dtype)
expected = torch.cat([
input @ lora_q.lora_a @ lora_q.lora_b * lora_q.scaling,
input @ lora_k.lora_a @ lora_k.lora_b * lora_k.scaling,
input @ lora_v.lora_a @ lora_v.lora_b * lora_v.scaling
],
dim=1)
lora_a_stacks = [
torch.zeros(8,
1,
lora_q.lora_a.shape[1],
lora_q.lora_a.shape[0],
device="cuda",
dtype=dtype)
] + [
torch.zeros(8,
1,
lora_k.lora_a.shape[1],
lora_k.lora_a.shape[0],
device="cuda",
dtype=dtype) for i in range(2)
]
lora_b_stacks = [
torch.zeros(8,
1,
lora_q.lora_b.shape[1],
lora_q.lora_b.shape[0],
device="cuda",
dtype=dtype)
] + [
torch.zeros(8,
1,
lora_k.lora_b.shape[1],
lora_k.lora_b.shape[0],
device="cuda",
dtype=dtype) for i in range(2)
]
for i in range(lora_a_stacks[0].shape[0]):
lora_a_stacks[0][i][0] = lora_q.lora_a.T
lora_b_stacks[0][i][0] = (lora_q.lora_b * lora_q.scaling).T
lora_a_stacks[1][i][0] = lora_k.lora_a.T
lora_b_stacks[1][i][0] = (lora_k.lora_b * lora_k.scaling).T
lora_a_stacks[2][i][0] = lora_v.lora_a.T
lora_b_stacks[2][i][0] = (lora_v.lora_b * lora_v.scaling).T
output = torch.zeros(k, sum(qkv), device="cuda", dtype=dtype)
_apply_lora_packed_nslice(
input, lora_a_stacks, lora_b_stacks,
torch.randint(0,
lora_a_stacks[0].shape[0], (len(input), ),
device="cuda"), output, (qkv[0], qkv[1], qkv[2]))
rtol, atol = TOLERANCES[dtype]
assert torch.allclose(expected, output, rtol=rtol, atol=atol)
output[:] = 0
_apply_lora_packed_nslice(input, lora_a_stacks, lora_b_stacks,
torch.full((len(input), ), -1, device="cuda"),
output, (qkv[0], qkv[1], qkv[2]))
assert torch.allclose(torch.zeros_like(output), output)
manager.reset_lora()
-258
View File
@@ -1,258 +0,0 @@
# Based on code from https://github.com/punica-ai/punica
import pytest
import torch
import vllm.lora.punica as punica
def assert_close(a, b):
rtol, atol = {
torch.float16: (5e-3, 5e-3),
torch.bfloat16: (3e-2, 2e-2),
torch.float32: (None, None),
}[a.dtype]
torch.testing.assert_close(a, b, rtol=rtol, atol=atol)
def _lora_ref_impl(
y_final: torch.Tensor,
x: torch.Tensor,
wa_T_all: torch.Tensor,
wb_T_all: torch.Tensor,
indicies: torch.LongTensor,
layer_idx: int,
scale: float,
):
y_stage_1 = torch.empty(
(x.size(0), wa_T_all.size(-2)),
dtype=torch.float32,
device=x.device,
)
bs = x.shape[0]
s = torch.tensor(scale, dtype=torch.float32, device=x.device)
for i, lora_idx in zip(range(bs), indicies.cpu().tolist()):
xi = x[i].unsqueeze(0).to(torch.float32)
wa = wa_T_all[lora_idx, layer_idx].transpose(-1, -2).to(torch.float32)
if wb_T_all is not None:
wb = wb_T_all[lora_idx, layer_idx].transpose(-1,
-2).to(torch.float32)
tmp = xi @ wa
y_stage_1[i] = tmp.squeeze(0)
y_final[i] += ((tmp @ wb).squeeze(0) *
s if wb_T_all is not None else y_stage_1[i])
return y_final, y_stage_1
H1 = H2 = [
128,
256,
512,
896,
1024,
1152,
1216,
1280,
1536,
1664,
2048,
2240,
2304,
2368,
2432,
2560,
2752,
3072,
3328,
3456,
3584,
3712,
4096,
4480,
4608,
4736,
4864,
5120,
5504,
5632,
5888,
6144,
6400,
6848,
6912,
7168,
7424,
8192,
8960,
9216,
9472,
10240,
11008,
11264,
13824,
14336,
14784,
14848,
15360,
18944,
22016,
22528,
24576,
27392,
27648,
29568,
29696,
32000,
32256,
32512,
32768,
33024,
36864,
43264,
49152,
49408,
60544,
60672,
64000,
64256,
102400,
102656,
128000,
128256,
]
H2 = [64] + H2
R = [1, 2, 4]
SEED = [0xabcdabcd987]
CUDA_DEVICES = [
f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
]
@pytest.mark.parametrize("dtype_str", ["float16", "bfloat16"])
@pytest.mark.parametrize("h1", H1)
@pytest.mark.parametrize("r", R)
@pytest.mark.parametrize("seed", SEED)
@torch.inference_mode()
def test_lora_a_extra_shapes(dtype_str, h1, r, seed):
torch.manual_seed(seed)
num_loras = 4
num_layers = 1
bs = 32
dtype = getattr(torch, dtype_str)
device = torch.device("cuda")
wa_T_all = torch.randn(num_loras,
num_layers,
r,
h1,
dtype=dtype,
device=device)
indices = torch.randint(num_loras, (bs, ), dtype=torch.long, device=device)
for layer_idx in range(num_layers):
x = torch.randn(bs, h1, dtype=dtype, device=device)
y = torch.randn(bs, r, dtype=dtype, device=device)
y_ref = y.clone()
_lora_ref_impl(
y_ref,
x,
wa_T_all,
None,
indices,
layer_idx,
1.0,
)
y_our = y.clone()
punica.bgmv(y_our, x, wa_T_all, indices, layer_idx, 1.0)
assert_close(y_ref, y_our)
@pytest.mark.parametrize("dtype_str", ["float16", "bfloat16"])
@pytest.mark.parametrize("h1", H1)
@pytest.mark.parametrize("h2", H2)
@pytest.mark.parametrize("seed", SEED)
@pytest.mark.parametrize("device", CUDA_DEVICES)
@torch.inference_mode()
def test_lora_correctness(dtype_str, h1, h2, seed, device):
torch.manual_seed(seed)
num_loras = 4
num_layers = 1
r = 8
bs = 32
scale = 0.123
dtype = getattr(torch, dtype_str)
torch.set_default_device(device)
wa_T_all = torch.randn(num_loras, num_layers, r, h1, dtype=dtype)
wb_T_all = torch.randn(num_loras, num_layers, h2, r, dtype=dtype)
indices = torch.randint(num_loras, (bs, ), dtype=torch.long)
for layer_idx in range(num_layers):
x = torch.randn(bs, h1, dtype=dtype)
y = torch.randn(bs, h2, dtype=dtype)
y_ref = y.clone()
_lora_ref_impl(y_ref, x, wa_T_all, wb_T_all, indices, layer_idx, scale)
y_our = y.clone()
punica.add_lora(y_our, x, wa_T_all, wb_T_all, indices, layer_idx,
scale)
assert_close(y_ref, y_our)
@pytest.mark.parametrize("dtype_str", ["float16", "bfloat16"])
@pytest.mark.parametrize("h1", H1)
@pytest.mark.parametrize("h2", H2)
@pytest.mark.parametrize("seed", SEED)
@pytest.mark.parametrize("device", CUDA_DEVICES)
@torch.inference_mode()
def test_lora_correctness_slice(dtype_str, h1, h2, seed, device):
if h2 % 3 != 0 or h2 // 3 not in H1:
pytest.skip("h2 must be divisible by 3 and in supported shapes")
torch.manual_seed(seed)
num_loras = 4
num_layers = 1
r = 8
bs = 32
scale = 0.123
dtype = getattr(torch, dtype_str)
torch.set_default_device(device)
wa_T_all_0 = torch.randn(num_loras, num_layers, r, h1, dtype=dtype)
wa_T_all_1 = torch.randn(num_loras, num_layers, r, h1, dtype=dtype)
wa_T_all_2 = torch.randn(num_loras, num_layers, r, h1, dtype=dtype)
wb_T_all_0 = torch.randn(num_loras, num_layers, h2 // 3, r, dtype=dtype)
wb_T_all_1 = torch.randn(num_loras, num_layers, h2 // 3, r, dtype=dtype)
wb_T_all_2 = torch.randn(num_loras, num_layers, h2 // 3, r, dtype=dtype)
indices = torch.randint(num_loras, (bs, ), dtype=torch.long)
for layer_idx in range(num_layers):
x = torch.randn(bs, h1, dtype=dtype)
y = torch.randn(bs, h2, dtype=dtype)
s = h2 // 3
y_ref = y.clone()
_lora_ref_impl(y_ref[:, :s], x, wa_T_all_0, wb_T_all_0, indices,
layer_idx, scale)
_lora_ref_impl(y_ref[:, s:s * 2], x, wa_T_all_1, wb_T_all_1, indices,
layer_idx, scale)
_lora_ref_impl(y_ref[:, s * 2:], x, wa_T_all_2, wb_T_all_2, indices,
layer_idx, scale)
y_our = y.clone()
punica.add_lora_slice(y_our, x, wa_T_all_0, wb_T_all_0, indices,
layer_idx, scale, 0, s)
punica.add_lora_slice(y_our, x, wa_T_all_1, wb_T_all_1, indices,
layer_idx, scale, s, s)
punica.add_lora_slice(y_our, x, wa_T_all_2, wb_T_all_2, indices,
layer_idx, scale, s * 2, s)
assert_close(y_ref[:, :s], y_our[:, :s])
assert_close(y_ref[:, s:s * 2], y_our[:, s:s * 2])
assert_close(y_ref[:, s * 2:], y_our[:, s * 2:])
+408
View File
@@ -0,0 +1,408 @@
"""
This script is mainly used to tests various hidden_sizes. We have collected the
hidden_sizes included in the LoRA models currently supported by vLLM. It tests
whether the corresponding Triton kernel can run normally when tensor parallelism
is set to [1, 2, 4, 8, 16, 32, 64].
"""
import random
from unittest.mock import patch
import pytest
import torch
from vllm.lora.ops.bgmv_expand import bgmv_expand
from vllm.lora.ops.bgmv_expand_slice import bgmv_expand_slice
from vllm.lora.ops.bgmv_shrink import bgmv_shrink
from vllm.lora.ops.sgmv_expand import sgmv_expand
from vllm.lora.ops.sgmv_expand_slice import sgmv_expand_slice
from vllm.lora.ops.sgmv_shrink import sgmv_shrink
from vllm.triton_utils.libentry import LibEntry
from .utils import (generate_data, generate_data_for_expand_nslices,
ref_torch_groupgemm)
HIDDEN_SIZES = [
128,
256,
512,
896,
1024,
1152,
1216,
1280,
1536,
1664,
2048,
2240,
2304,
2368,
2432,
2560,
2752,
3072,
3328,
3456,
3584,
3712,
4096,
4480,
4608,
4736,
4864,
5120,
5504,
5632,
5888,
6144,
6400,
6848,
6912,
7168,
7424,
8192,
8960,
9216,
9472,
10240,
11008,
11264,
13824,
14336,
14784,
14848,
15360,
18944,
22016,
22528,
24576,
27392,
27648,
29568,
29696,
32000,
32256,
32512,
32768,
33024,
36864,
43264,
49152,
49408,
60544,
60672,
64000,
64256,
102400,
102656,
128000,
128256,
]
#The size of TP
divisibility = [1, 2, 4, 8, 16, 32, 64]
all_hidden_size = []
for div in divisibility:
for hidden_size in HIDDEN_SIZES:
all_hidden_size.append(hidden_size // div)
HIDDEN_SIZES = list(set(all_hidden_size))
BATCHES = [4]
NUM_LORA = [4]
DTYPES = [torch.float16, torch.bfloat16]
MAX_RANKS = [32]
SCALES = [0.5]
SEED = [0]
CUDA_DEVICES = [f"cuda:{0}"]
def assert_close(a, b):
rtol, atol = {
torch.float16: (6e-2, 6e-2),
torch.bfloat16: (6e-2, 6e-2),
torch.float32: (1e-2, 1e-2),
}[a.dtype]
torch.testing.assert_close(a, b, rtol=rtol, atol=atol)
@pytest.mark.parametrize("batches", BATCHES)
@pytest.mark.parametrize("num_loras", NUM_LORA)
@pytest.mark.parametrize("rank", MAX_RANKS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("scaling", SCALES)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("op_type", ["shrink", "expand"])
@pytest.mark.parametrize("seed", SEED)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_punica_sgmv(
batches: int,
num_loras: int,
rank: int,
hidden_size: int,
scaling: float,
dtype: torch.dtype,
op_type: str,
seed: int,
device: str,
):
random.seed(seed)
torch.set_default_device(device)
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
seq_length = 128
(
inputs_tensor,
lora_weights,
our_out_tensor,
ref_out_tensor,
b_seq_start_loc,
lora_indices_tensor,
seq_len_tensor,
indices,
) = generate_data(
batches,
hidden_size,
num_loras,
rank,
seq_length,
dtype,
op_type,
device,
)
max_seq_length = seq_len_tensor.max()
if isinstance(max_seq_length, tuple):
max_seq_length = max_seq_length[0].item()
else:
max_seq_length = max_seq_length.item()
if op_type == "shrink":
sgmv_shrink(
inputs_tensor,
lora_weights,
our_out_tensor,
b_seq_start_loc,
seq_len_tensor,
lora_indices_tensor,
batches,
max_seq_length,
scaling,
)
else:
sgmv_expand(
inputs_tensor,
lora_weights,
our_out_tensor,
b_seq_start_loc,
seq_len_tensor,
lora_indices_tensor,
batches,
max_seq_length,
add_inputs=True,
)
ref_torch_groupgemm(
ref_out_tensor,
inputs_tensor,
lora_weights,
lora_indices_tensor,
seq_len_tensor,
batches,
scaling if op_type == "shrink" else 1.0,
op_type,
)
if op_type == "shrink":
ref_out_tensor = ref_out_tensor.to(torch.float32)
assert_close(our_out_tensor, ref_out_tensor)
@pytest.mark.parametrize("batches", BATCHES)
@pytest.mark.parametrize("num_loras", NUM_LORA)
@pytest.mark.parametrize("rank", MAX_RANKS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("scaling", SCALES)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("op_type", ["shrink", "expand"])
@pytest.mark.parametrize("seed", SEED)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_punica_bgmv(
batches: int,
num_loras: int,
rank: int,
hidden_size: int,
scaling: float,
dtype: torch.dtype,
op_type: str,
seed: int,
device: str,
):
from vllm.lora.ops.bgmv_expand import _bgmv_expand_kernel
from vllm.lora.ops.bgmv_shrink import _bgmv_shrink_kernel
random.seed(seed)
torch.set_default_device(device)
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
seq_length = 1
(
inputs_tensor,
lora_weights,
our_out_tensor,
ref_out_tensor,
b_seq_start_loc,
lora_indices_tensor,
seq_len_tensor,
indices,
) = generate_data(
batches,
hidden_size,
num_loras,
rank,
seq_length,
dtype,
op_type,
device,
)
if op_type == "shrink":
# The current _bgmv_shrink_kernel does not require the libentry
# decoration. The purpose of adding this patch is to test the
# correctness of libentry.
with patch(
"vllm.lora.ops.bgmv_shrink._bgmv_shrink_kernel",
LibEntry(_bgmv_shrink_kernel),
):
bgmv_shrink(
inputs_tensor,
lora_weights,
our_out_tensor,
indices,
scaling,
)
else:
# ditto
with patch(
"vllm.lora.ops.bgmv_expand._bgmv_expand_kernel",
LibEntry(_bgmv_expand_kernel),
):
bgmv_expand(
inputs_tensor,
lora_weights,
our_out_tensor,
indices,
add_inputs=True,
)
ref_torch_groupgemm(
ref_out_tensor,
inputs_tensor,
lora_weights,
lora_indices_tensor,
seq_len_tensor,
batches,
scaling if op_type == "shrink" else 1.0,
op_type,
)
if op_type == "shrink":
ref_out_tensor = ref_out_tensor.to(torch.float32)
assert_close(our_out_tensor, ref_out_tensor)
@pytest.mark.parametrize("batches", BATCHES)
@pytest.mark.parametrize("num_loras", NUM_LORA)
@pytest.mark.parametrize("rank", MAX_RANKS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("nslices", [2, 3])
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("op_type", ["sgmv", "bgmv"])
@pytest.mark.parametrize("seed", SEED)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_punica_expand_nslices(
batches: int,
num_loras: int,
rank: int,
hidden_size: int,
nslices: int,
dtype: torch.dtype,
op_type: str,
seed: int,
device: str,
):
from vllm.lora.ops.bgmv_expand_slice import _bgmv_expand_slice_kernel
random.seed(seed)
torch.set_default_device(device)
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
seq_length = 128 if op_type == "sgmv" else 1
(
inputs_tensor,
lora_weights_lst,
our_outputs,
ref_outputs,
b_seq_start_loc,
lora_indices_tensor,
seq_len_tensor,
indices,
) = generate_data_for_expand_nslices(
batches,
hidden_size,
num_loras,
rank,
seq_length,
dtype,
nslices,
device,
)
max_seq_length = seq_len_tensor.max()
if isinstance(max_seq_length, tuple):
max_seq_length = max_seq_length[0].item()
else:
max_seq_length = max_seq_length.item()
slice_offset = 0
for index in range(nslices):
lora_weights = lora_weights_lst[index]
if op_type == "sgmv":
sgmv_expand_slice(
inputs_tensor,
lora_weights,
our_outputs,
b_seq_start_loc,
seq_len_tensor,
lora_indices_tensor,
batches,
max_seq_length,
slice_offset,
hidden_size,
add_inputs=True,
)
else:
# The current _bgmv_expand_slice_kernel does not require the
# libentry decoration. The purpose of adding this patch is to test
# the correctness of libentry.
with patch(
"vllm.lora.ops.bgmv_expand_slice._bgmv_expand_slice_kernel",
LibEntry(_bgmv_expand_slice_kernel),
):
bgmv_expand_slice(
inputs_tensor,
lora_weights,
our_outputs,
indices,
slice_offset,
slice_size=hidden_size,
add_inputs=True,
)
ref_torch_groupgemm(
ref_outputs[:, slice_offset:slice_offset + hidden_size],
inputs_tensor,
lora_weights,
lora_indices_tensor,
seq_len_tensor,
batches,
1.0,
op_type="expand",
)
slice_offset += hidden_size
assert_close(our_outputs, ref_outputs)
+342
View File
@@ -0,0 +1,342 @@
"""
This script is mainly used to test whether trtion kernels can run normally
under different conditions, including various batches, numbers of LoRA , and
maximum ranks.
"""
import random
from unittest.mock import patch
import pytest
import torch
from vllm.lora.ops.bgmv_expand import bgmv_expand
from vllm.lora.ops.bgmv_expand_slice import bgmv_expand_slice
from vllm.lora.ops.bgmv_shrink import bgmv_shrink
from vllm.lora.ops.sgmv_expand import sgmv_expand
from vllm.lora.ops.sgmv_expand_slice import sgmv_expand_slice
from vllm.lora.ops.sgmv_shrink import sgmv_shrink
from vllm.triton_utils.libentry import LibEntry
from .utils import (generate_data, generate_data_for_expand_nslices,
ref_torch_groupgemm)
HIDDEN_SIZES = [3424, 4096, 4097]
BATCHES = [1, 4, 16, 32]
NUM_LORA = [1, 4, 8, 16, 32, 64, 128]
DTYPES = [torch.float16, torch.bfloat16]
MAX_RANKS = [1, 4, 8, 16, 32, 64, 128]
SCALES = [0.5]
SEED = [0]
CUDA_DEVICES = [f"cuda:{0}"]
def assert_close(a, b):
rtol, atol = {
torch.float16: (6e-2, 6e-2),
torch.bfloat16: (6e-2, 6e-2),
torch.float32: (1e-2, 1e-2),
}[a.dtype]
torch.testing.assert_close(a, b, rtol=rtol, atol=atol)
@pytest.mark.parametrize("batches", BATCHES)
@pytest.mark.parametrize("num_loras", NUM_LORA)
@pytest.mark.parametrize("rank", MAX_RANKS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("scaling", SCALES)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("op_type", ["shrink", "expand"])
@pytest.mark.parametrize("seed", SEED)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_punica_sgmv(
batches: int,
num_loras: int,
rank: int,
hidden_size: int,
scaling: float,
dtype: torch.dtype,
op_type: str,
seed: int,
device: str,
):
random.seed(seed)
torch.set_default_device(device)
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
seq_length = 128
(
inputs_tensor,
lora_weights,
our_out_tensor,
ref_out_tensor,
b_seq_start_loc,
lora_indices_tensor,
seq_len_tensor,
indices,
) = generate_data(
batches,
hidden_size,
num_loras,
rank,
seq_length,
dtype,
op_type,
device,
)
max_seq_length = seq_len_tensor.max()
if isinstance(max_seq_length, tuple):
max_seq_length = max_seq_length[0].item()
else:
max_seq_length = max_seq_length.item()
if op_type == "shrink":
sgmv_shrink(
inputs_tensor,
lora_weights,
our_out_tensor,
b_seq_start_loc,
seq_len_tensor,
lora_indices_tensor,
batches,
max_seq_length,
scaling,
)
else:
sgmv_expand(
inputs_tensor,
lora_weights,
our_out_tensor,
b_seq_start_loc,
seq_len_tensor,
lora_indices_tensor,
batches,
max_seq_length,
add_inputs=True,
)
ref_torch_groupgemm(
ref_out_tensor,
inputs_tensor,
lora_weights,
lora_indices_tensor,
seq_len_tensor,
batches,
scaling if op_type == "shrink" else 1.0,
op_type,
)
if op_type == "shrink":
ref_out_tensor = ref_out_tensor.to(torch.float32)
assert_close(our_out_tensor, ref_out_tensor)
@pytest.mark.parametrize("batches", BATCHES)
@pytest.mark.parametrize("num_loras", NUM_LORA)
@pytest.mark.parametrize("rank", MAX_RANKS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("scaling", SCALES)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("op_type", ["shrink", "expand"])
@pytest.mark.parametrize("seed", SEED)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_punica_bgmv(
batches: int,
num_loras: int,
rank: int,
hidden_size: int,
scaling: float,
dtype: torch.dtype,
op_type: str,
seed: int,
device: str,
):
from vllm.lora.ops.bgmv_expand import _bgmv_expand_kernel
from vllm.lora.ops.bgmv_shrink import _bgmv_shrink_kernel
random.seed(seed)
torch.set_default_device(device)
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
seq_length = 1
(
inputs_tensor,
lora_weights,
our_out_tensor,
ref_out_tensor,
b_seq_start_loc,
lora_indices_tensor,
seq_len_tensor,
indices,
) = generate_data(
batches,
hidden_size,
num_loras,
rank,
seq_length,
dtype,
op_type,
device,
)
if op_type == "shrink":
# The current _bgmv_shrink_kernel does not require the libentry
# decoration. The purpose of adding this patch is to test the
# correctness of libentry.
with patch(
"vllm.lora.ops.bgmv_shrink._bgmv_shrink_kernel",
LibEntry(_bgmv_shrink_kernel),
):
bgmv_shrink(
inputs_tensor,
lora_weights,
our_out_tensor,
indices,
scaling,
)
else:
# ditto
with patch(
"vllm.lora.ops.bgmv_expand._bgmv_expand_kernel",
LibEntry(_bgmv_expand_kernel),
):
bgmv_expand(
inputs_tensor,
lora_weights,
our_out_tensor,
indices,
add_inputs=True,
)
ref_torch_groupgemm(
ref_out_tensor,
inputs_tensor,
lora_weights,
lora_indices_tensor,
seq_len_tensor,
batches,
scaling if op_type == "shrink" else 1.0,
op_type,
)
if op_type == "shrink":
ref_out_tensor = ref_out_tensor.to(torch.float32)
assert_close(our_out_tensor, ref_out_tensor)
@pytest.mark.parametrize("batches", BATCHES)
@pytest.mark.parametrize("num_loras", NUM_LORA)
@pytest.mark.parametrize("rank", MAX_RANKS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("nslices", [2, 3])
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("op_type", ["sgmv", "bgmv"])
@pytest.mark.parametrize("seed", SEED)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_punica_expand_nslices(
batches: int,
num_loras: int,
rank: int,
hidden_size: int,
nslices: int,
dtype: torch.dtype,
op_type: str,
seed: int,
device: str,
):
from vllm.lora.ops.bgmv_expand_slice import _bgmv_expand_slice_kernel
random.seed(seed)
torch.set_default_device(device)
torch.random.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
seq_length = 128 if op_type == "sgmv" else 1
(
inputs_tensor,
lora_weights_lst,
our_outputs,
ref_outputs,
b_seq_start_loc,
lora_indices_tensor,
seq_len_tensor,
indices,
) = generate_data_for_expand_nslices(
batches,
hidden_size,
num_loras,
rank,
seq_length,
dtype,
nslices,
device,
)
max_seq_length = seq_len_tensor.max()
if isinstance(max_seq_length, tuple):
max_seq_length = max_seq_length[0].item()
else:
max_seq_length = max_seq_length.item()
slice_offset = 0
for index in range(nslices):
lora_weights = lora_weights_lst[index]
if op_type == "sgmv":
sgmv_expand_slice(
inputs_tensor,
lora_weights,
our_outputs,
b_seq_start_loc,
seq_len_tensor,
lora_indices_tensor,
batches,
max_seq_length,
slice_offset,
hidden_size,
add_inputs=True,
)
else:
# The current _bgmv_expand_slice_kernel does not require the
# libentry decoration. The purpose of adding this patch is to test
# the correctness of libentry.
with patch(
"vllm.lora.ops.bgmv_expand_slice._bgmv_expand_slice_kernel",
LibEntry(_bgmv_expand_slice_kernel),
):
bgmv_expand_slice(
inputs_tensor,
lora_weights,
our_outputs,
indices,
slice_offset,
slice_size=hidden_size,
add_inputs=True,
)
ref_torch_groupgemm(
ref_outputs[:, slice_offset:slice_offset + hidden_size],
inputs_tensor,
lora_weights,
lora_indices_tensor,
seq_len_tensor,
batches,
1.0,
op_type="expand",
)
slice_offset += hidden_size
assert_close(our_outputs, ref_outputs)
if __name__ == "__main__":
from itertools import product
lst = list(
product(
BATCHES,
NUM_LORA,
MAX_RANKS,
[1.0],
[torch.float16],
["expand"],
SEED,
CUDA_DEVICES,
))
for ele in lst:
test_punica_bgmv(*ele)
print(f"{ele},pass")
+27 -21
View File
@@ -64,14 +64,16 @@ def test_quant_model_lora(tinyllama_lora_files, model, tp_size):
# if torch.cuda.device_count() < tp_size:
# pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
llm = vllm.LLM(model=model.model_path,
enable_lora=True,
max_num_seqs=16,
max_loras=4,
max_model_len=400,
tensor_parallel_size=tp_size,
quantization=model.quantization,
trust_remote_code=True)
llm = vllm.LLM(
model=model.model_path,
enable_lora=True,
max_num_seqs=16,
max_loras=4,
max_model_len=400,
tensor_parallel_size=tp_size,
gpu_memory_utilization=0.2, #avoid OOM
quantization=model.quantization,
trust_remote_code=True)
if model.quantization is None:
expected_no_lora_output = [
@@ -156,24 +158,28 @@ def test_quant_model_tp_equality(tinyllama_lora_files, model):
# if torch.cuda.device_count() < 2:
# pytest.skip(f"Not enough GPUs for tensor parallelism {2}")
llm_tp1 = vllm.LLM(model=model.model_path,
enable_lora=True,
max_num_seqs=16,
max_loras=4,
tensor_parallel_size=1,
quantization=model.quantization,
trust_remote_code=True)
llm_tp1 = vllm.LLM(
model=model.model_path,
enable_lora=True,
max_num_seqs=16,
max_loras=4,
tensor_parallel_size=1,
gpu_memory_utilization=0.2, #avoid OOM
quantization=model.quantization,
trust_remote_code=True)
output_tp1 = do_sample(llm_tp1, tinyllama_lora_files, lora_id=1)
del llm_tp1
cleanup()
llm_tp2 = vllm.LLM(model=model.model_path,
enable_lora=True,
max_num_seqs=16,
max_loras=4,
tensor_parallel_size=2,
quantization=model.quantization)
llm_tp2 = vllm.LLM(
model=model.model_path,
enable_lora=True,
max_num_seqs=16,
max_loras=4,
tensor_parallel_size=2,
gpu_memory_utilization=0.2, #avoid OOM
quantization=model.quantization)
output_tp2 = do_sample(llm_tp2, tinyllama_lora_files, lora_id=1)
del llm_tp2
+148
View File
@@ -86,3 +86,151 @@ class DummyLoRAManager:
packed_lora = PackedLoRALayerWeights.pack(base_loras)
self.set_module_lora(module_name, packed_lora)
return packed_lora
def assert_close(a, b):
rtol, atol = {
torch.float16: (6e-2, 6e-2),
torch.bfloat16: (6e-2, 6e-2),
torch.float32: (1e-2, 1e-2),
}[a.dtype]
torch.testing.assert_close(a, b, rtol=rtol, atol=atol)
def ref_torch_groupgemm(
out_tensor,
inputs,
lora_weights,
lora_indices_tensor,
seq_len_tensor,
batches,
scaling,
op_type,
) -> torch.Tensor:
out_list = []
current_offset = 0
for lora_index, b_length in zip(range(batches), seq_len_tensor):
input_weight = inputs[current_offset:b_length + current_offset, :]
current_offset += b_length
lora_weight = lora_weights[lora_indices_tensor[lora_index]]
result = torch.nn.functional.linear(input_weight, lora_weight)
result *= scaling
out_list.append(result)
cat_result = torch.cat(out_list, dim=0)
if op_type == "expand":
out_tensor += cat_result
else:
out_tensor.copy_(cat_result)
return
def generate_data(batches, hidden_size, lora_nums, max_rank, seq_length, dtype,
op_type, device):
seq_len_tensor = torch.randint(seq_length, seq_length + 1,
(batches, )).to(device)
b_seq_start_loc = torch.cumsum(
torch.tensor([0] + seq_len_tensor[:-1].tolist(), dtype=torch.long),
dim=0,
).to(device)
total_tokens = seq_len_tensor.sum()
if op_type == "shrink":
inputs_tensor = torch.rand((total_tokens, hidden_size),
dtype=dtype).to(device)
lora_weights = torch.rand(
(lora_nums, max_rank, hidden_size), # col-major
dtype=dtype,
).to(device)
# shrink op need atomic_add, so output is initinized by 0
ref_out_tensor = torch.zeros((total_tokens, max_rank),
dtype=dtype,
device=inputs_tensor.device)
# NOTE shrink kernel using torch.float32 as output type
our_out_tensor = torch.zeros((total_tokens, max_rank),
dtype=torch.float32).to(device)
else:
inputs_tensor = torch.rand(
(total_tokens, max_rank),
dtype=dtype,
).to(device)
lora_weights = torch.rand(
(lora_nums, hidden_size, max_rank), # col-major
dtype=dtype,
).to(device)
# expand op needs to complete y+=a@lora_b, so output is
# initinized randomly
ref_out_tensor = torch.rand(
(total_tokens, hidden_size),
dtype=dtype,
).to(device)
# Ensure the same input.
our_out_tensor = ref_out_tensor.clone()
lora_indices_tensor = torch.randint(0,
lora_nums - 1 if lora_nums > 1 else 1,
(batches, )).to(device)
indices = torch.zeros((total_tokens), dtype=torch.long).to(device)
current_offset = 0
for b_id in range(batches):
lora_index = lora_indices_tensor[b_id]
indices[current_offset:current_offset +
seq_len_tensor[b_id]].copy_(lora_index)
current_offset += seq_len_tensor[b_id].item()
return (
inputs_tensor,
lora_weights,
our_out_tensor,
ref_out_tensor,
b_seq_start_loc,
lora_indices_tensor,
seq_len_tensor,
indices,
)
def generate_data_for_expand_nslices(batches, hidden_size, lora_nums, max_rank,
seq_length, dtype, nslices, device):
seq_len_tensor = torch.randint(seq_length, seq_length + 1,
(batches, )).to(device)
b_seq_start_loc = torch.cumsum(
torch.tensor([0] + seq_len_tensor[:-1].tolist(), dtype=torch.long),
dim=0,
).to(device)
total_tokens = seq_len_tensor.sum()
inputs_tensor = torch.rand(
(total_tokens, max_rank),
dtype=dtype,
).to(device)
lora_weights_lst = []
for _ in range(nslices):
lora_weights_lst.append(
torch.rand(
(lora_nums, hidden_size, max_rank), # col-major
dtype=dtype,
).to(device))
# expand op needs to complete y+=a@lora_b, so output is
# initinized randomly
ref_out_tensor = torch.rand((total_tokens, hidden_size * nslices),
dtype=dtype).to(device)
# Ensure the same input.
our_out_tensor = ref_out_tensor.clone()
lora_indices_tensor = torch.randint(0,
lora_nums - 1 if lora_nums > 1 else 1,
(batches, ))
indices = torch.zeros((total_tokens), dtype=torch.long).to(device)
current_offset = 0
for b_id in range(batches):
lora_index = lora_indices_tensor[b_id]
indices[current_offset:current_offset +
seq_len_tensor[b_id]] = lora_index.item()
current_offset += seq_len_tensor[b_id].item()
lora_indices_tensor = lora_indices_tensor.to(device)
return (
inputs_tensor,
lora_weights_lst,
our_out_tensor,
ref_out_tensor,
b_seq_start_loc,
lora_indices_tensor,
seq_len_tensor,
indices,
)