Merge branch 'v0.6.2-eval' into v0.6.2-dev

# Conflicts:
#	csrc/attention/static_switch_tc.h
#	vllm/model_executor/layers/vocab_parallel_embedding.py
#	vllm/model_executor/model_loader/utils.py
#	vllm/model_executor/models/llama.py

benchmarks/benchmark_prefix_caching.py

@@ -35,6 +35,8 @@ from transformers import PreTrainedTokenizerBase
 from vllm import LLM, SamplingParams
 from vllm.utils import FlexibleArgumentParser
 
+from triton.common.backend import compute_core_version_key
+
 try:
     from vllm.transformers_utils.tokenizer import get_tokenizer
 except ImportError:
@@ -44,6 +46,7 @@ PROMPT = "You are a helpful assistant in recognizes the content of tables in mar
 
 
 def test_prefix(llm=None, sampling_params=None, prompts=None):
+    version_key = compute_core_version_key()
     start_time = time.time()
 
     llm.generate(prompts, sampling_params=sampling_params)

csrc/attention/static_switch_tc.h

@@ -43,6 +43,9 @@
     } else if (HEADDIM == 160) {           \
       constexpr static int HEAD_SIZE = 160; \
       return __VA_ARGS__();                \
+    } else if (HEADDIM == 192) {           \
+      constexpr static int HEAD_SIZE = 192; \
+      return __VA_ARGS__();                \
     } else if (HEADDIM == 256) {           \
       constexpr static int HEAD_SIZE = 256; \
       return __VA_ARGS__();                \

vllm/model_executor/layers/vocab_parallel_embedding.py

@@ -22,7 +22,6 @@ class UnquantizedEmbeddingMethod(QuantizeMethodBase):
     
     def __init__(self):
         self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
-        self.use_lm_nn = os.environ.get('LM_NN') == '1'
 
     def create_weights(self, layer: torch.nn.Module,
                        input_size_per_partition: int,
@@ -42,7 +41,7 @@ class UnquantizedEmbeddingMethod(QuantizeMethodBase):
               layer: torch.nn.Module,
               x: torch.Tensor,
               bias: Optional[torch.Tensor] = None) -> torch.Tensor:
-        if self.use_llama_nn and self.use_lm_nn:
+        if self.use_llama_nn and os.environ['LM_NN'] == '1':
             if bias is not None:
                 if len(x.shape) == 2: 
                     return torch.addmm(bias, x, layer.weight)

vllm/model_executor/models/__init__.py

@@ -219,4 +219,4 @@ class ModelRegistry:
 
 __all__ = [
     "ModelRegistry",
-]
+]
\ No newline at end of file

vllm/model_executor/models/chatglm.py

@@ -726,4 +726,4 @@ class ChatGLMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
                     ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                     weight.data.copy_(_weight)
                     
-                    weight.data=weight.data.reshape(ori_shape[1], -1)
+                    weight.data=weight.data.reshape(ori_shape[1], -1)
\ No newline at end of file

vllm/model_executor/models/llama.py

@@ -574,18 +574,20 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA):
                 "self_attn.qkv_proj.weight",
                 "self_attn.o_proj.weight",
                 "mlp.gate_up_proj.weight",
-                "mlp.down_proj.weight"
+                "mlp.down_proj.weight",
+                "lm_head.weight"
             ]
             
-            if self.use_lm_nn:
-                lay_key_words.append("lm_head.weight")
-                
             combined_words = "|".join(lay_key_words)
             
             lay_qkv_words = ["self_attn.qkv_proj.weight"]   
             qkv_words = "|".join(lay_qkv_words)          
             
             for layername, weight in params_dict.items():
+                if "lm_head.weight" in layername:
+                    os.environ['LM_NN'] = '1'  
+                else:
+                    os.environ['LM_NN'] = '0' 
                 matches = re.findall(combined_words, layername)
                 if matches:         
                     if self.use_gemm_pad and gemm_bank_conf(weight.data.shape[0]):