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, 00:45 UTC
[Misc] Update w2 scale loading for GPTQMarlinMoE (#12757)
Dipika Sikka committed
Feb 6, 2025 at 04:02 UTC
7ca9934fe773edf8680aed287b0a05cb195bd8e4
3 files changed
+21
-8
tests/weight_loading/models-large.txt
+2
@@ -1,5 +1,7 @@
1
compressed-tensors, nm-testing/Mixtral-8x7B-Instruct-v0.1-W4A16-quantized, main
2
compressed-tensors, nm-testing/Mixtral-8x7B-Instruct-v0.1-W4A16-channel-quantized, main
3
compressed-tensors, nm-testing/Mixtral-8x7B-Instruct-v0.1-W8A16-quantized, main
4
+compressed-tensors, nm-testing/test-w4a16-mixtral-actorder-group, main
5
gptq_marlin, TheBloke/Mixtral-8x7B-v0.1-GPTQ, main
6
+gptq_marlin, TheBloke/Mixtral-8x7B-v0.1-GPTQ, gptq-8bit-128g-actorder_True
7
awq_marlin, casperhansen/deepseek-coder-v2-instruct-awq, main
\ No newline at end of file
vllm/model_executor/layers/fused_moe/layer.py
+2
-2
@@ -302,8 +302,8 @@ class FusedMoE(torch.nn.Module):
302
"weight_loader": self.weight_loader,
303
}
304
# need full intermediate size pre-sharding for WNA16 act order
305
- if (self.quant_method.__class__.__name__ ==
306
- "CompressedTensorsWNA16MoEMethod"):
305
+ if (self.quant_method.__class__.__name__
306
+ in ("GPTQMarlinMoEMethod", "CompressedTensorsWNA16MoEMethod")):
307
moe_quant_params["intermediate_size_full"] = intermediate_size
308
309
self.quant_method.create_weights(layer=self, **moe_quant_params)
vllm/model_executor/layers/quantization/gptq_marlin.py
+17
-6
@@ -323,13 +323,18 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
323
params_dtype: torch.dtype,
324
**extra_weight_attrs,
325
):
326
- # Currently assuming is_k_full is always True
327
- # (input size per partition is the same as full input size)
328
- # Supports only sym for now (no zp)
326
+ intermediate_size_full = extra_weight_attrs.pop(
327
+ "intermediate_size_full")
328
+
329
+ self.is_k_full = (not self.quant_config.desc_act) or (
330
+ intermediate_size_per_partition == intermediate_size_full)
331
+
332
if self.quant_config.group_size != -1:
333
scales_size13 = hidden_size // self.quant_config.group_size
331
- scales_size2 = (intermediate_size_per_partition //
332
- self.quant_config.group_size)
334
+ w2_scales_size = (intermediate_size_full
335
+ if self.quant_config.desc_act else
336
+ intermediate_size_per_partition)
337
+ scales_size2 = (w2_scales_size // self.quant_config.group_size)
338
strategy = FusedMoeWeightScaleSupported.GROUP.value
339
else:
340
scales_size13 = 1
@@ -385,6 +390,9 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
390
)
391
layer.register_parameter("w2_scales", w2_scales)
392
set_weight_attrs(w2_scales, extra_weight_attrs)
393
+ # dont shard the w2 scales when running act order
394
+ set_weight_attrs(w2_scales,
395
+ {"load_full_w2": self.quant_config.desc_act})
396
# up_proj scales
397
w13_qzeros = torch.nn.Parameter(
398
torch.empty(num_experts,
@@ -406,6 +414,9 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
414
)
415
layer.register_parameter("w2_qzeros", w2_qzeros)
416
set_weight_attrs(w2_qzeros, extra_weight_attrs)
417
+ # dont shard the w2 scales when running act order
418
+ set_weight_attrs(w2_qzeros,
419
+ {"load_full_w2": self.quant_config.desc_act})
420
w13_g_idx = torch.nn.Parameter(
421
torch.empty(
422
num_experts,
@@ -575,4 +586,4 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
586
sort_indices1=layer.w13_g_idx_sort_indices,
587
sort_indices2=layer.w2_g_idx_sort_indices,
588
num_bits=self.quant_config.quant_type.size_bits,
578
- ).to(orig_dtype)
589
+ is_k_full=self.is_k_full).to(orig_dtype)