x
This commit is contained in:
+10
-1
@@ -114,7 +114,16 @@ def launch_model(model_id, config, model_path, gpu_count):
|
||||
# Get configuration
|
||||
valid_tp = config.get("valid_tp", [1])
|
||||
max_tp = max(valid_tp) if valid_tp else 1
|
||||
tp_size = min(gpu_count, max_tp)
|
||||
|
||||
# Check for manual TP_SIZE override
|
||||
tp_size_env = os.getenv("TP_SIZE")
|
||||
if tp_size_env:
|
||||
tp_size = int(tp_size_env)
|
||||
log(f"TP_SIZE environment variable set: {tp_size}")
|
||||
if tp_size not in valid_tp:
|
||||
log(f"WARNING: TP_SIZE={tp_size} is not in valid_tp={valid_tp}, proceeding anyway")
|
||||
else:
|
||||
tp_size = min(gpu_count, max_tp)
|
||||
|
||||
ctx = int(config.get("ctx", 8192))
|
||||
max_seqs = int(config.get("max_num_seqs", 64))
|
||||
|
||||
Reference in New Issue
Block a user