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 15, 14:45 UTC
Enable PTPC FP8 for CompressedTensorsW8A8Fp8MoEMethod (triton fused_moe) (#16537)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Apr 13, 2025 at 08:55 UTC
d085a4408244502dd496cae0ccbc524352e87ef8
1 file changed
+68
-39
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
+68
-39
@@ -88,14 +88,23 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
88
self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
89
"input_activations")
90
91
- if not (self.weight_quant.strategy == QuantizationStrategy.TENSOR
92
- and self.input_quant.strategy == QuantizationStrategy.TENSOR):
91
+ per_tensor = (self.weight_quant.strategy == QuantizationStrategy.TENSOR
92
+ and self.input_quant.strategy
93
+ == QuantizationStrategy.TENSOR)
94
+ per_channel = (
95
+ self.weight_quant.strategy == QuantizationStrategy.CHANNEL
96
+ and self.input_quant.strategy == QuantizationStrategy.TOKEN)
97
+ if not (per_tensor or per_channel):
98
raise ValueError(
94
- "For FP8 Fused MoE layers, only per-tensor scales "
95
- "for weights and activations are supported. Found "
99
+ "For FP8 Fused MoE layers, we require per tensor "
100
+ "or channelwise, dynamic per token quantization. Found "
101
f"{self.weight_quant}, {self.input_quant}")
102
103
self.static_input_scales = not self.input_quant.dynamic
104
+ if self.static_input_scales and per_channel:
105
+ raise ValueError(
106
+ "For FP8 Fused MoE layer, we require either per tensor or "
107
+ "channelwise, dynamic per token quantization.")
108
109
def create_weights(self, layer: torch.nn.Module, num_experts: int,
110
hidden_size: int, intermediate_size_per_partition: int,
@@ -123,24 +132,40 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
132
set_weight_attrs(w2_weight, extra_weight_attrs)
133
134
# WEIGHT_SCALES
126
- # Allocate 2 scales for w1 and w3 respectively.
127
- # They will be combined to a single scale after weight loading.
128
- w13_weight_scale = torch.nn.Parameter(torch.ones(num_experts,
129
- 2,
130
- dtype=torch.float32),
131
- requires_grad=False)
132
- layer.register_parameter("w13_weight_scale", w13_weight_scale)
135
+ if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
136
+ # Allocate 2 scales for w1 and w3 respectively.
137
+ # They are combined to a single scale after weight loading.
138
+ w13_weight_scale = torch.nn.Parameter(torch.ones(
139
+ num_experts, 2, dtype=torch.float32),
140
+ requires_grad=False)
141
+ layer.register_parameter("w13_weight_scale", w13_weight_scale)
142
+ w2_weight_scale = torch.nn.Parameter(torch.ones(
143
+ num_experts, dtype=torch.float32),
144
+ requires_grad=False)
145
+ layer.register_parameter("w2_weight_scale", w2_weight_scale)
146
+ # Add PER-TENSOR quantization for FusedMoE.weight_loader.
147
+ extra_weight_attrs.update(
148
+ {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
149
+ set_weight_attrs(w13_weight_scale, extra_weight_attrs)
150
+ set_weight_attrs(w2_weight_scale, extra_weight_attrs)
151
134
- w2_weight_scale = torch.nn.Parameter(torch.ones(num_experts,
135
- dtype=torch.float32),
136
- requires_grad=False)
137
- layer.register_parameter("w2_weight_scale", w2_weight_scale)
138
- # Add the quantization method used (per tensor/grouped/channel)
139
- # to ensure the weight scales are loaded in properly
140
- extra_weight_attrs.update(
141
- {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
142
- set_weight_attrs(w13_weight_scale, extra_weight_attrs)
143
- set_weight_attrs(w2_weight_scale, extra_weight_attrs)
152
+ elif self.weight_quant.strategy == QuantizationStrategy.CHANNEL:
153
+ w13_weight_scale = torch.nn.Parameter(torch.ones(
154
+ num_experts,
155
+ 2 * intermediate_size_per_partition,
156
+ 1,
157
+ dtype=torch.float32),
158
+ requires_grad=False)
159
+ layer.register_parameter("w13_weight_scale", w13_weight_scale)
160
+ w2_weight_scale = torch.nn.Parameter(torch.ones(
161
+ num_experts, hidden_size, 1, dtype=torch.float32),
162
+ requires_grad=False)
163
+ layer.register_parameter("w2_weight_scale", w2_weight_scale)
164
+ # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
165
+ extra_weight_attrs.update(
166
+ {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value})
167
+ set_weight_attrs(w13_weight_scale, extra_weight_attrs)
168
+ set_weight_attrs(w2_weight_scale, extra_weight_attrs)
169
170
# INPUT_SCALES
171
if self.static_input_scales:
@@ -163,6 +188,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
188
# Fp8 moe kernels require a single activation scale.
189
# We take the max of all the scales in case they differ.
190
if self.static_input_scales:
191
+ assert self.input_quant.strategy == QuantizationStrategy.TENSOR
192
if (layer.w13_input_scale is None or layer.w2_input_scale is None):
193
raise ValueError(
194
"QuantConfig has static quantization, but found "
@@ -204,24 +230,25 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
230
layer.w2_input_scale = torch.nn.Parameter(w2_input_scale,
231
requires_grad=False)
232
207
- # Fp8 moe kernel needs single weight scale for w13 per expert.
208
- # We take the max then dequant and requant each expert.
209
- assert layer.w13_weight_scale is not None
210
- shard_size = layer.intermediate_size_per_partition
211
- max_w13_scales = layer.w13_weight_scale.max(dim=1).values
212
- for expert_id in range(layer.local_num_experts):
213
- start = 0
214
- for shard_id in range(2):
215
- dq_weight = per_tensor_dequantize(
216
- layer.w13_weight[expert_id][start:start + shard_size, :],
217
- layer.w13_weight_scale[expert_id][shard_id])
218
- layer.w13_weight[expert_id][
219
- start:start + shard_size, :], _ = ops.scaled_fp8_quant(
220
- dq_weight, max_w13_scales[expert_id])
221
- start += shard_size
222
-
223
- layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
224
- requires_grad=False)
233
+ # For Per-TENSOR case, Fp8 moe kernel needs single weight scale
234
+ # for w13 per expert. Use max then dequant and requant each expert.
235
+ if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
236
+ assert layer.w13_weight_scale is not None
237
+ shard_size = layer.intermediate_size_per_partition
238
+ max_w13_scales = layer.w13_weight_scale.max(dim=1).values
239
+ for expert_id in range(layer.local_num_experts):
240
+ start = 0
241
+ for shard_id in range(2):
242
+ dq_weight = per_tensor_dequantize(
243
+ layer.w13_weight[expert_id][start:start +
244
+ shard_size, :],
245
+ layer.w13_weight_scale[expert_id][shard_id])
246
+ layer.w13_weight[expert_id][
247
+ start:start + shard_size, :], _ = ops.scaled_fp8_quant(
248
+ dq_weight, max_w13_scales[expert_id])
249
+ start += shard_size
250
+ layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
251
+ requires_grad=False)
252
253
def apply(
254
self,
@@ -265,6 +292,8 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
292
activation=activation,
293
apply_router_weight_on_input=apply_router_weight_on_input,
294
use_fp8_w8a8=True,
295
+ per_channel_quant=self.weight_quant.strategy ==
296
+ QuantizationStrategy.CHANNEL,
297
global_num_experts=global_num_experts,
298
expert_map=expert_map,
299
w1_scale=layer.w13_weight_scale,