p2o6e100
/

ASVDPhi2-2.5B-v1-q0f16-MLC

Model card Files Files and versions Community

张瀚灵 commited on Oct 22, 2024

Commit

f416b0e

·

1 Parent(s): 12d5e7e

update config

Files changed (1) hide show

mlc-chat-config.json +3 -3

mlc-chat-config.json CHANGED Viewed

@@ -12,7 +12,7 @@
     "position_embedding_base": 10000.0,
     "partial_rotary_factor": 0.4,
     "num_key_value_heads": 32,
-    "context_window_size": 2048,
     "prefill_chunk_size": 128,
     "head_dim": 80,
     "truncation_ranks": {
@@ -302,12 +302,12 @@
       ]
     },
     "tensor_parallel_shards": 1,
-    "max_batch_size": 128
   },
   "vocab_size": 51200,
   "context_window_size": 2048,
   "sliding_window_size": -1,
-  "prefill_chunk_size": 128,
   "attention_sink_size": -1,
   "tensor_parallel_shards": 1,
   "pipeline_parallel_stages": 1,

     "position_embedding_base": 10000.0,
     "partial_rotary_factor": 0.4,
     "num_key_value_heads": 32,
+    "context_window_size": 128,
     "prefill_chunk_size": 128,
     "head_dim": 80,
     "truncation_ranks": {
       ]
     },
     "tensor_parallel_shards": 1,
+    "max_batch_size": 1
   },
   "vocab_size": 51200,
   "context_window_size": 2048,
   "sliding_window_size": -1,
+  "prefill_chunk_size": 2048,
   "attention_sink_size": -1,
   "tensor_parallel_shards": 1,
   "pipeline_parallel_stages": 1,