Merge branch 'v0.9.2-dev-fix-DS-unquant-rebase' into 'v0.9.2-dev'

fix: 细化量化模型开启融合场景

See merge request dcutoolkit/deeplearing/vllm!292

vllm/model_executor/layers/linear.py

@@ -33,18 +33,8 @@ from vllm.platforms import current_platform
 
 import os
 from vllm.model_executor.utils import gemm_bank_conf
-
-if envs.USE_FUSED_RMS_QUANT:
-    try:
-        from lmslim.quantize.quant_ops import lm_faster_rmsquant
-    except Exception as e:
-        print(f"Error: Import fused rmsquant error: {e}") 
-if envs.USE_FUSED_SILU_MUL_QUANT:        
-    try:
-        # from lightop import fuse_silu_mul_quant
-        from lmslim.quantize.quant_ops import lm_fuse_silu_mul_quant
-    except Exception as e:
-        print(f"Error: Import fused silu_mul_qunat error: {e}")
+from lmslim.quantize.quant_ops import lm_faster_rmsquant
+from lmslim.quantize.quant_ops import lm_fuse_silu_mul_quant
         
 logger = init_logger(__name__)
 

vllm/model_executor/model_loader/utils.py

@@ -261,7 +261,7 @@ def get_model_architecture(
                 #     os.environ['VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT'] = '1'
                 if not envs.is_set("VLLM_SCHED_ENABLE_MINIMAL_INJECTION"):
                     os.environ['VLLM_SCHED_ENABLE_MINIMAL_INJECTION'] = '1'
-                if model_config.quantization is not None:
+                if model_config.quantization in {"slimquant_w4a8", "slimquant_w4a8_marlin", "slimquant_compressed_tensors_marlin", "compressed-tensors"}:
                     if not envs.is_set("USE_FUSED_RMS_QUANT"):
                         os.environ['USE_FUSED_RMS_QUANT'] = '1'
                     if not envs.is_set("USE_FUSED_SILU_MUL_QUANT"):
@@ -306,7 +306,8 @@ def get_model_architecture(
                 #     os.environ['VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT'] = '1'
                 if not envs.is_set("VLLM_SCHED_ENABLE_MINIMAL_INJECTION"):
                     os.environ['VLLM_SCHED_ENABLE_MINIMAL_INJECTION'] = '1'
-                if model_config.quantization is not None:
+                if model_config.quantization in {"slimquant_w4a8", "slimquant_w4a8_marlin", "slimquant_compressed_tensors_marlin", "compressed-tensors"}:
+                    
                     if not envs.is_set("USE_FUSED_RMS_QUANT"):
                         os.environ['USE_FUSED_RMS_QUANT'] = '1'
                     if not envs.is_set("USE_FUSED_SILU_MUL_QUANT"):