added context_length field

Signed-off-by: Zhuoran-Liu <[email protected]>

Author: WarningRan

infer/vllm/process

@@ -57,14 +57,42 @@ def extract_model_version(parsed_model, input_model):
     return None
 
 def determine_precision(model_name, default_precision):
-    """
-    Determine precision based on model name.
-    If model name contains 'fp8' (case insensitive), return 'fp8', otherwise return default.
-    """
+
     if model_name and 'fp8' in model_name.lower():
         return 'fp8'
     return default_precision
 
+def extract_context_length_direct(log_driver):
+
+    if not os.path.exists(log_driver):
+        return None
+    
+    try:
+        for line in open(log_driver, errors='ignore'):
+            if line.startswith('FMWORK ARG') and '--engine:max_model_len@' in line:
+                match = re.search(r'--engine:max_model_len@\s*(\d+)', line)
+                if match:
+                    return int(match.group(1))
+    except Exception:
+        pass
+    
+    return None
+
+def extract_context_length_server(cmd_server):
+
+    if not os.path.exists(cmd_server):
+        return None
+    
+    try:
+        content = open(cmd_server).read()
+        match = re.search(r'--max-model-len\s+(\d+)', content)
+        if match:
+            return int(match.group(1))
+    except Exception:
+        pass
+    
+    return None
+
 def get_server_completion_info(args_path):
 
     cmd_client = os.path.join(args_path, 'client.cmd')
@@ -234,17 +262,26 @@ def process_direct(args):
                 model_version = None
                 final_model_name = model_name
 
+            # Determine precision based on model name
+            final_precision = determine_precision(model_name, args.precision)
+
+            # Determine context length
+            context_length = extract_context_length_direct(log_driver)
+            if context_length is None and input_size is not None and output_size is not None:
+                context_length = input_size + output_size
+
             hits.append({
                 'timestamp'     : time_start,
                 'metadata_id'   : args.metadata_id,
                 'engine'        : 'fmwork/infer/vllm',
                 'model'         : final_model_name,
                 'model_version' : model_version,
-                'precision'     : args.precision,
+                'precision'     : final_precision,
                 'input'         : input_size,
                 'output'        : output_size,
                 'batch'         : batch_size,
                 'tp'            : tp_size,
+                'context_length': context_length,
                 'opts'          : opts,
                 'warmup'        : round(warmup, 3) if warmup is not None else None,
                 'setup'         : setup,
@@ -276,17 +313,26 @@ def process_direct(args):
             model_version = None
             final_model_name = model_name
 
+        # Determine precision based on model name
+        final_precision = determine_precision(model_name, args.precision)
+        
+        # Determine context length
+        context_length = extract_context_length_direct(log_driver)
+        if context_length is None and input_size is not None and output_size is not None:
+            context_length = input_size + output_size
+        
         hits.append({
             'timestamp'     : None,
             'metadata_id'   : args.metadata_id,
             'engine'        : 'fmwork/infer/vllm',
             'model'         : final_model_name,
             'model_version' : model_version,
-            'precision'     : args.precision,
+            'precision'     : final_precision,
             'input'         : input_size,
             'output'        : output_size,
             'batch'         : batch_size,
             'tp'            : tp_size,
+            'context_length': context_length,
             'opts'          : opts,
             'warmup'        : None,
             'setup'         : setup,
@@ -519,6 +565,11 @@ def process_server(args):
     # Determine precision based on model name
     final_precision = determine_precision(model_name, args.precision)
 
+    # Determine context length
+    context_length = extract_context_length_server(cmd_server)
+    if context_length is None and input_size is not None and output_size is not None:
+        context_length = input_size + output_size
+
     hits = [{
         'timestamp'     : time_start,
         'metadata_id'   : args.metadata_id,
@@ -530,6 +581,7 @@ def process_server(args):
         'output'        : output_size,
         'batch'         : batch_size,
         'tp'            : tp_size,
+        'context_length': context_length,
         'opts'          : opts,
         'warmup'        : None,
         'setup'         : None,