fix

Author: hiworldwzj

lightllm/distributed/communication_op.py

@@ -36,6 +36,7 @@
 )
 from lightllm.utils.device_utils import get_device_sm_count
 from lightllm.utils.sgl_utils import HAS_SGL_KERNEL
+from lightllm.utils.light_utils import HAS_LIGHTLLM_KERNEL
 from contextlib import nullcontext, contextmanager
 
 logger = init_logger(__name__)
@@ -56,9 +57,6 @@
     HAS_DEEPEP = False
     logger.info("deep_ep is not installed, you can't use the api of it.")
 
-# TODO: lightllm_kernel release.
-HAS_LIGHTLLM_KERNEL = False
-
 
 class CustomProcessGroup:
     def __init__(self):

lightllm/utils/vllm_utils.py

@@ -14,6 +14,7 @@
 except:
     HAS_VLLM = False
     cutlass_scaled_mm = None
+    vllm_ops = None
     logger.warning(
         "vllm is not installed, you can't use the api of it. \
                    You can solve it by running `pip install vllm`."

test/model/model_infer.py

@@ -377,7 +377,7 @@ def tppart_model_infer(args, model_kvargs, batch_size, input_len, output_len, an
                         b_ready_cache_len=b_ready_cache_len,
                         is_prefill=True,
                     ),
-                    log_dir=f"./logs_sglang_4k/forward_prefill_{model_kvargs['rank_id']}",
+                    log_dir=f"./logs/forward_prefill_{model_kvargs['rank_id']}",
                 )
         except Exception as e:
             print(str(e))
@@ -417,7 +417,7 @@ def tppart_model_infer(args, model_kvargs, batch_size, input_len, output_len, an
                         b_seq_len,
                         total_token_num,
                     ),
-                    log_dir=f"./logs_sglang_4k/forward_decode_{model_kvargs['rank_id']}",
+                    log_dir=f"./logs/forward_decode_{model_kvargs['rank_id']}",
                 )
         else:
             logits = decode(
@@ -442,7 +442,7 @@ def tppart_model_infer(args, model_kvargs, batch_size, input_len, output_len, an
                         b_seq_len,
                         total_token_num,
                     ),
-                    log_dir=f"./logs_sglang_4k/forward_decode_{model_kvargs['rank_id']}",
+                    log_dir=f"./logs/forward_decode_{model_kvargs['rank_id']}",
                 )
 
         prob_out = torch.softmax(logits, dim=-1)