update bits (#986)

Signed-off-by: He, Xin3 <xin3.he@intel.com>

Author: xin3he

test/test_cuda/test_auto_scheme.py

@@ -84,7 +84,7 @@ def test_shared_layers(self):
     @multi_card
     def test_multi_card(self):
         model_name = "/models/Qwen3-0.6B"
-        target_bits = 5.265
+        target_bits = 4.5
         for device_map in ["auto", "0,1", "0", None]:
             scheme = AutoScheme(avg_bits=target_bits, options=("NVFP4"))
             ar = AutoRound(model=model_name, scheme=scheme, iters=0, nsamples=1, device_map=device_map)
@@ -96,7 +96,7 @@ def test_multi_card(self):
     @multi_card
     def test_multi_card_1(self):
         model_name = "/models/Qwen3-0.6B"
-        target_bits = 5.265
+        target_bits = 4.5
         from transformers import AutoModelForCausalLM, AutoTokenizer
 
         tokenizer = AutoTokenizer.from_pretrained(model_name)
@@ -110,7 +110,7 @@ def test_multi_card_1(self):
 
     def test_non_low_gpu_mem_usage(self):
         model_name = "/models/Qwen3-0.6B"
-        target_bits = 5.265
+        target_bits = 4.5
         # for device_map in ["auto", "0,1", "0", None]:
         scheme = AutoScheme(avg_bits=target_bits, options=("NVFP4"), low_gpu_mem_usage=False, device_map="auto")
 
@@ -123,7 +123,7 @@ def test_non_low_gpu_mem_usage(self):
     @multi_card
     def test_dict_device_map(self):
         model_name = "/models/Qwen3-8B"
-        target_bits = 8.755
+        target_bits = 8.25
         device_map = {"up_proj": 0, "down_proj": 1}
 
         scheme = AutoScheme(avg_bits=target_bits, options=("MXFP8"))