Update to transformers 4.35 and update some rust TLS dependencies

Also updated patched transformers files with upstream updates

Author: njhill

Cargo.lock

@@ -24,10 +24,9 @@ moka = { version = "0.11.3", features = ["future"] }
 nohash-hasher = "^0.2.0"
 num = "^0.4.1"
 hyper = "^0.14.27" # Override to address CVE-2023-26964
-openssl = "^0.10.55" # Override to address WS-2023-0082, WS-2023-0083, WS-2023-0195
-openssl-sys = "^0.9.90" # Override to address WS-2023-0082, WS-2023-0083, WS-2023-0195
-rustls-webpki = "^0.100.2" # Override to address WS-2023-0305, CVE-2018-16875
-parking_lot = "^0.12.1"
+openssl = "^0.10.59" # Override to address WS-2023-0082, WS-2023-0083, WS-2023-0195
+openssl-sys = "^0.9.95" # Override to address WS-2023-0082, WS-2023-0083, WS-2023-0195
+rustls-webpki = "^0.101.7" # Override to address WS-2023-0305, CVE-2018-16875
 rand = "^0.8.5"
 serde = "^1.0.173"
 serde_json = "^1.0.103"
@@ -38,7 +37,7 @@ thiserror = "^1.0.50"
 tokenizers = "0.14.1"
 tokio = { version = "1.33.0", features = ["rt", "rt-multi-thread", "parking_lot", "signal", "sync", "fs"] }
 tokio-rustls = "^0.24.1"
-rustls = "0.21.5"
+rustls = "0.21.8"
 tracing = "^0.1.40"
 tracing-subscriber = { version = "0.3.17", features = ["json"] }
 prost = "^0.12.1"

router/Cargo.toml

@@ -21,7 +21,7 @@ safetensors = "^0.4.0"
 sentencepiece = "^0.1.99"
 datasets = { version = "^2.14.6", optional = true }
 texttable = { version = "^1.7.0", optional = true }
-transformers = "4.34.1"
+transformers = "4.35.0"
 optimum = { version = "1.13.2", extras = ["onnxruntime-gpu"], optional = true }
 onnxruntime = { version = "1.16.1", optional = true }
 onnxruntime-gpu = { version = "1.16.1", optional = true }

server/poetry.lock

@@ -1002,6 +1002,11 @@ def forward(
 
 
 class T5ForConditionalGeneration(T5PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [
+        "decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
+    ]
+    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
+
     def __init__(self, config: T5Config, weights):
         super().__init__(config)
         self.model_dim = config.d_model
@@ -1160,9 +1165,18 @@ def prepare_inputs_for_generation(
         encoder_outputs=None,
         **kwargs,
     ):
-        # cut decoder_input_ids if past is used
+        # cut decoder_input_ids if past_key_values is used
         if past_key_values is not None:
-            input_ids = input_ids[:, -1:]
+            past_length = past_key_values[0][0].shape[2]
+
+            # Some generation methods already pass only the last input ID
+            if input_ids.shape[1] > past_length:
+                remove_prefix_length = past_length
+            else:
+                # Default to old behavior: keep only final ID
+                remove_prefix_length = input_ids.shape[1] - 1
+
+            input_ids = input_ids[:, remove_prefix_length:]
 
         return {
             "decoder_input_ids": input_ids,