biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernel][Attention] Separate Attention.kv_scale into k_scale and v_scale (#6081)

Browse Source
This commit is contained in:
Michael Goin
2024-07-16 18:31:32 -04:00
committed by GitHub
parent 160e1d8c99
commit 978aed5300
33 changed files with 317 additions and 185 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
vllm/model_executor/layers/quantization/fp8.py
View File

@@ -407,31 +407,56 @@ class Fp8KVCacheMethod(QuantizeMethodBase):
self.quant_config = quant_config
def create_weights(self, layer: torch.nn.Module):
"""Create "weight" (aka kv_scale) for an attention layer.
"""Create "weight" (aka k_scale and v_scale) for an attention layer.
Args:
layer: The layer that is using the QuantizeMethodBase factory.
"""
# Initialize the KV cache scale to 1.0 as the default value.
# If the kv_scale appears in the checkpoint, it will be
# Initialize the KV cache scales to -1.0, which is an invalid value.
# If the k/v_scale appears in the checkpoint, it will be
# overwritten when loading weights.
layer.kv_scale = Parameter(torch.tensor(1.0), requires_grad=False)
layer.k_scale = Parameter(torch.tensor(-1.0), requires_grad=False)
layer.v_scale = Parameter(torch.tensor(-1.0), requires_grad=False)
def apply(self, layer: torch.nn.Module) -> torch.Tensor:
raise RuntimeError("Fp8KVCacheMethod.apply should not be called.")
def process_weights_after_loading(self, layer: Module) -> None:
# If the kv-cache dtype is auto, we enforce the kv-scale to be 1.0
# If the kv-cache dtype is auto, we enforce the k/v_scale to be 1.0
# regardless whether the kv-scale is available in the checkpoint.
if layer.kv_cache_dtype != "auto":
kv_scale = layer.kv_scale.to("cpu").tolist()
if not isinstance(kv_scale, float):
if layer.k_scale > 0.0 and layer.v_scale > 0.0:
# We prefer to use separate k_scale and v_scale if present
k_scale = layer.k_scale.to("cpu").tolist()
v_scale = layer.v_scale.to("cpu").tolist()
elif layer.k_scale < 0.0 and layer.v_scale < 0.0:
# If no scales were loaded (both scales are invalid negative
# values), use the default value of 1.0
k_scale = Parameter(torch.tensor(1.0), requires_grad=False)
v_scale = Parameter(torch.tensor(1.0), requires_grad=False)
else:
# If we find a single kv_scale in the checkpoint, we remap
# kv_scale to k_scale during weight loading, and duplicate
# k_scale to v_scale here
assert layer.k_scale > 0.0
scale_to_duplicate = max(layer.k_scale, layer.v_scale)
k_scale = scale_to_duplicate.to("cpu").tolist()
v_scale = scale_to_duplicate.to("cpu").tolist()
if not isinstance(k_scale, float) or not isinstance(
v_scale, float):
raise ValueError("Only support per-tensor scaling factor "
"for fp8 KV cache")
layer._kv_scale = kv_scale
if layer._kv_scale == 1.0 and "e5m2" not in layer.kv_cache_dtype:
# These are used in the final Attention.forward()
layer._k_scale = k_scale
layer._v_scale = v_scale
if (layer._k_scale == 1.0 and layer._v_scale == 1.0
and "e5m2" not in layer.kv_cache_dtype):
print_warning_once(
"Using KV cache scaling factor 1.0 for fp8_e4m3. This may "
"cause accuracy issues. Please make sure kv-cache scaling "
"factor is available in the fp8 checkpoint.")
del layer.kv_scale
"Using KV cache scaling factor 1.0 for fp8_e4m3. This "
"may cause accuracy issues. Please make sure k/v_scale "
"scaling factors are available in the fp8 checkpoint.")
del layer.k_scale
del layer.v_scale