[ Misc ] Refactor Marlin Python Utilities (#6082)

Co-authored-by: Robert Shaw <rshaw@neuralmagic.com>
This commit is contained in:
Robert Shaw
2024-07-11 15:40:11 +00:00
committed by GitHub
co-authored by Robert Shaw
parent 55f692b46e
commit b675069d74
12 changed files with 690 additions and 728 deletions
+25 -21
View File
@@ -5,19 +5,21 @@ Run `pytest tests/kernels/marlin/test_marlin_gemm.py`.
import pytest
import torch
from tests.quantization.utils import is_quant_method_supported
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.gptq_marlin import (
GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
GPTQ_MARLIN_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_SUPPORTED_NUM_BITS,
marlin_permute_scales)
from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_NUM_BITS)
from vllm.model_executor.layers.quantization.utils.marlin_perms import (
marlin_perm)
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
MarlinWorkspace, compute_max_diff, is_marlin_supported, marlin_24_quantize,
marlin_quantize, marlin_weights, pack_fp8_to_int32)
GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
GPTQ_MARLIN_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_SUPPORTED_NUM_BITS,
marlin_permute_scales)
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
pack_fp8_to_int32)
from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
MarlinWorkspace, get_weight_perm, marlin_quantize, marlin_weights)
from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
marlin_24_quantize)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
gptq_pack, quantize_weights, sort_weights)
@@ -42,11 +44,16 @@ MNK_FACTORS = [
DTYPES = [torch.float16, torch.bfloat16]
def compute_max_diff(output, output_ref):
return torch.mean(torch.abs(output - output_ref)) / torch.mean(
torch.abs(output_ref))
def rand_data(shape, dtype=torch.float16):
return torch.randn(shape, dtype=dtype, device="cuda")
@pytest.mark.skipif(not is_marlin_supported(),
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
reason="Marlin is not supported on this GPU type.")
@pytest.mark.parametrize("k_chunk", MARLIN_K_CHUNKS)
@pytest.mark.parametrize("n_chunk", MARLIN_N_CHUNKS)
@@ -93,8 +100,8 @@ def test_marlin_repack(k_chunk, n_chunk, num_bits, group_size, act_order,
q_w, g_idx, sort_indices = sort_weights(q_w, g_idx)
# Pack to Marlin format
marlin_q_w_1 = marlin_weights(q_w, size_k, size_n, num_bits,
marlin_perm[num_bits])
weight_perm = get_weight_perm(num_bits)
marlin_q_w_1 = marlin_weights(q_w, size_k, size_n, num_bits, weight_perm)
# Run Marlin repack GPU kernel
marlin_q_w_2 = ops.gptq_marlin_repack(
@@ -109,7 +116,7 @@ def test_marlin_repack(k_chunk, n_chunk, num_bits, group_size, act_order,
assert torch.allclose(marlin_q_w_1, marlin_q_w_2)
@pytest.mark.skipif(not is_marlin_supported(),
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
reason="Marlin is not supported on this GPU type.")
@pytest.mark.parametrize("k_chunk", MARLIN_K_CHUNKS)
@pytest.mark.parametrize("n_chunk", MARLIN_N_CHUNKS)
@@ -174,7 +181,7 @@ def test_marlin_gemm(
assert max_diff < 0.04
@pytest.mark.skipif(not is_marlin_supported(),
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
reason="Marlin is not supported on this GPU type.")
@pytest.mark.parametrize("k_chunk", MARLIN_24_K_CHUNKS)
@pytest.mark.parametrize("n_chunk", MARLIN_24_N_CHUNKS)
@@ -222,7 +229,7 @@ def test_marlin_24_gemm(k_chunk, n_chunk, num_bits, group_size, mnk_factors):
assert max_diff < 0.04
@pytest.mark.skipif(not is_marlin_supported(),
@pytest.mark.skipif(not is_quant_method_supported("fp8"),
reason="Marlin is not supported on this GPU type.")
@pytest.mark.parametrize("k_chunk", MARLIN_K_CHUNKS)
@pytest.mark.parametrize("n_chunk", MARLIN_N_CHUNKS)
@@ -268,13 +275,10 @@ def test_fp8_marlin_gemm(
# expand it to channelwise
scales = weight_scale.repeat(1, size_n).to(a_input.dtype).to("cuda")
# Permute scales
marlin_scales = marlin_permute_scales(
s=scales,
size_k=size_k,
size_n=size_n,
group_size=-1,
num_bits=8,
)
marlin_scales = marlin_permute_scales(s=scales,
size_k=size_k,
size_n=size_n,
group_size=-1)
workspace = MarlinWorkspace(size_n, GPTQ_MARLIN_MIN_THREAD_N,
GPTQ_MARLIN_MAX_PARALLEL)