Upload folder using huggingface_hub

decoder/config.json

@@ -68,10 +68,10 @@
         "max_input_len": 14,
         "max_seq_len": 460,
         "opt_batch_size": null,
-        "max_batch_size": 32,
-        "max_beam_width": 1,
-        "max_num_tokens": 8192,
-        "opt_num_tokens": 32,
+        "max_batch_size": 16,
+        "max_beam_width": 5,
+        "max_num_tokens": 7360,
+        "opt_num_tokens": 80,
         "max_prompt_embedding_table_size": 0,
         "kv_cache_type": "PAGED",
         "gather_context_logits": false,

decoder/rank0.engine

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7fcc78c4ea30b236c27d3a08c9ab4df8a25e1ca4f8b8ad900390ab861cb39e61
-size 1118724028
+oid sha256:07787f0a6dd56bdd45bc26c13dc64f15f627ed561dd64afbf48c14c941e646ab
+size 1118703372

encoder/config.json

@@ -46,10 +46,10 @@
         "max_input_len": 1024,
         "max_seq_len": 2048,
         "opt_batch_size": null,
-        "max_batch_size": 32,
+        "max_batch_size": 16,
         "max_beam_width": 1,
         "max_num_tokens": 8192,
-        "opt_num_tokens": 32,
+        "opt_num_tokens": 16,
         "max_prompt_embedding_table_size": 0,
         "kv_cache_type": "PAGED",
         "gather_context_logits": false,

encoder/rank0.engine

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c6f2d23587e6daf6d5525abf8b74b51346278395a5726d9fdfa4090c4553c3ee
-size 651985812
+oid sha256:7a63a150c0f14186904d9726e9cc37e436a018520f1abd7e8c462531d6df5a39
+size 652001460