Recalc scales from user (vllm-project#774)

mul scale input in factor = 448/240

---------

Co-authored-by: Michał Kuligowski <[email protected]>

Author: linoybu

requirements-hpu.txt

@@ -8,4 +8,4 @@ pandas
 tabulate
 setuptools>=61
 setuptools-scm>=8
-vllm-hpu-extension @ git+https://github.com/HabanaAI/vllm-hpu-extension.git@8087a98
+vllm-hpu-extension @ git+https://github.com/HabanaAI/vllm-hpu-extension.git@bb47de4

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py

@@ -5,6 +5,7 @@
 import torch
 from compressed_tensors.quantization import QuantizationStrategy
 from torch.nn import Parameter
+from vllm_hpu_extension.ops import get_hpu_gaudi2_scale_factor, is_hpu_gaudi2
 
 from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
     CompressedTensorsScheme)
@@ -84,8 +85,10 @@ def process_weights_after_loading(self, layer) -> None:
 
         # INPUT SCALE
         if self.is_static_input_scheme and hasattr(layer, 'input_scale'):
-            layer.input_scale = Parameter(layer.input_scale.max(),
-                                          requires_grad=False)
+            input_scale = layer.input_scale.max()
+            if is_hpu_gaudi2():
+                input_scale = input_scale * get_hpu_gaudi2_scale_factor()
+            layer.input_scale = Parameter(input_scale, requires_grad=False)
         else:
             layer.input_scale = None
 

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -3,6 +3,7 @@
 from typing import List, Optional, Tuple, Union
 
 import torch
+from vllm_hpu_extension.ops import get_hpu_gaudi2_scale_factor, is_hpu_gaudi2
 
 from vllm import _custom_ops as ops
 from vllm.platforms import current_platform
@@ -101,10 +102,8 @@ def requantize_with_max_scale(
         logical_widths: List[int]) -> Tuple[torch.Tensor, torch.Tensor]:
     # Max scale to be used for requanitzation.
     max_w_scale = weight_scale.max()
-    if current_platform.is_hpu() and htexp._get_device_type(
-    ) == htexp.synDeviceType.synDeviceGaudi2:
-        max_w_scale = max_w_scale * (torch.finfo(torch.float8_e4m3fn).max /
-                                     torch.finfo(torch.float8_e4m3fnuz).max)
+    if is_hpu_gaudi2():
+        max_w_scale = max_w_scale * get_hpu_gaudi2_scale_factor()
     # QKV / MLP is fused in the on disk checkpoint if any of the
     # weight scales are still set to the default since we initialize
     # N weight scales for N shards but we only load 1 weight scale