A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 19:30 UTC
[Bugfix][FP8] Fix dynamic FP8 Marlin quantization (#7219)
Michael Goin committed
Aug 7, 2024 at 14:23 UTC
5223199e03ac3729eb60043a1ef57156c8af1bc9
3 files changed
+33
-5
tests/quantization/test_fp8.py
+15
-4
@@ -9,6 +9,7 @@ from tests.quantization.utils import is_quant_method_supported
9
from vllm import _custom_ops as ops
10
from vllm.model_executor.layers.quantization.fp8 import (Fp8KVCacheMethod,
11
Fp8LinearMethod)
12
+from vllm.platforms import current_platform
13
14
MODELS = [
15
"neuralmagic/Meta-Llama-3-8B-Instruct-FP8-KV",
@@ -20,7 +21,12 @@ MODELS = [
21
@pytest.mark.skipif(not is_quant_method_supported("fp8"),
22
reason="FP8 is not supported on this GPU type.")
23
@pytest.mark.parametrize("model_id", MODELS)
23
-def test_model_load_and_run(vllm_runner, model_id: str):
24
+@pytest.mark.parametrize("force_marlin", [False, True])
25
+def test_model_load_and_run(vllm_runner, model_id: str, force_marlin: bool,
26
+ monkeypatch) -> None:
27
+ if force_marlin:
28
+ monkeypatch.setenv("VLLM_TEST_FORCE_FP8_MARLIN", "1")
29
+
30
with vllm_runner(model_id) as llm:
31
# note: this does not test accuracy, just that we can run through
32
# see lm-eval tests for accuracy
@@ -61,7 +67,12 @@ def test_kv_cache_model_load_and_run(vllm_runner, model_id: str):
67
@pytest.mark.skipif(not is_quant_method_supported("fp8"),
68
reason="FP8 is not supported on this GPU type.")
69
@pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8"])
64
-def test_load_fp16_model(vllm_runner, kv_cache_dtype: str) -> None:
70
+@pytest.mark.parametrize("force_marlin", [False, True])
71
+def test_load_fp16_model(vllm_runner, kv_cache_dtype: str, force_marlin: bool,
72
+ monkeypatch) -> None:
73
+ if force_marlin:
74
+ monkeypatch.setenv("VLLM_TEST_FORCE_FP8_MARLIN", "1")
75
+
76
with vllm_runner("facebook/opt-125m",
77
quantization="fp8",
78
kv_cache_dtype=kv_cache_dtype) as llm:
@@ -75,9 +86,9 @@ def test_load_fp16_model(vllm_runner, kv_cache_dtype: str) -> None:
86
assert attn._k_scale == 1.0
87
assert attn._v_scale == 1.0
88
78
- capability = torch.cuda.get_device_capability()
89
+ capability = current_platform.get_device_capability()
90
capability = capability[0] * 10 + capability[1]
80
- if capability >= 89:
91
+ if capability >= 89 and not force_marlin:
92
# For GPUs with hardware support, we keep weights in fp8
93
assert fc1.weight.dtype == torch.float8_e4m3fn
94
else:
vllm/envs.py
+8
@@ -52,6 +52,7 @@ if TYPE_CHECKING:
52
CMAKE_BUILD_TYPE: Optional[str] = None
53
VERBOSE: bool = False
54
VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
55
+ VLLM_TEST_FORCE_FP8_MARLIN: bool = False
56
57
58
def get_default_cache_root():
@@ -342,6 +343,13 @@ environment_variables: Dict[str, Callable[[], Any]] = {
343
lambda:
344
(os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower() in
345
("1", "true")),
346
+
347
+ # If set, forces FP8 Marlin to be used for FP8 quantization regardless
348
+ # of the hardware support for FP8 compute.
349
+ "VLLM_TEST_FORCE_FP8_MARLIN":
350
+ lambda:
351
+ (os.environ.get("VLLM_TEST_FORCE_FP8_MARLIN", "0").strip().lower() in
352
+ ("1", "true")),
353
}
354
355
# end-env-vars-definition
vllm/model_executor/layers/quantization/fp8.py
+10
-1
@@ -4,6 +4,7 @@ import torch
4
from torch.nn import Module
5
from torch.nn.parameter import Parameter
6
7
+import vllm.envs as envs
8
from vllm import _custom_ops as ops
9
from vllm.logger import init_logger
10
from vllm.model_executor.layers.fused_moe import FusedMoE, FusedMoEMethodBase
@@ -118,7 +119,7 @@ class Fp8LinearMethod(LinearMethodBase):
119
# kernel for fast weight-only FP8 quantization
120
capability = current_platform.get_device_capability()
121
capability = capability[0] * 10 + capability[1]
121
- self.use_marlin = capability < 89
122
+ self.use_marlin = capability < 89 or envs.VLLM_TEST_FORCE_FP8_MARLIN
123
124
def create_weights(
125
self,
@@ -174,6 +175,14 @@ class Fp8LinearMethod(LinearMethodBase):
175
qweight, weight_scale = ops.scaled_fp8_quant(layer.weight,
176
scale=None)
177
178
+ # If using marlin (w8a16), kernel uses channelwise weights,
179
+ # so extend the weight scales to be channelwise.
180
+ if self.use_marlin:
181
+ assert weight_scale.numel() == 1
182
+ weight_scale = convert_to_channelwise(
183
+ weight_scale.expand(len(layer.logical_widths)),
184
+ layer.logical_widths)
185
+
186
# Update the layer with the new values.
187
layer.weight = Parameter(qweight.t(), requires_grad=False)
188
layer.weight_scale = Parameter(weight_scale, requires_grad=False)