v1.3.1

Cargo.toml

@@ -8,7 +8,7 @@ members = [
 ]
 
 [workspace.package]
-version = "1.3.0"
+version = "1.3.1"
 edition = "2021"
 authors = ["Olivier Dehaene"]
 homepage = "https://github.com/huggingface/text-generation-inference"

docs/openapi.json

@@ -10,7 +10,7 @@
       "name": "Apache 2.0",
       "url": "https://www.apache.org/licenses/LICENSE-2.0"
     },
-    "version": "1.3.0"
+    "version": "1.3.1"
   },
   "paths": {
     "/": {

integration-tests/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "text-generation-integration-tests"
-version = "1.3.0"
+version = "1.3.1"
 description = "Text Generation Inference integration tests"
 authors = ["Nicolas Patry <[email protected]>"]
 

server/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "text-generation-server"
-version = "1.3.0"
+version = "1.3.1"
 description = "Text Generation Inference Python gRPC Server"
 authors = ["Olivier Dehaene <[email protected]>"]
 

server/text_generation_server/models/custom_modeling/flash_mistral_modeling.py

@@ -391,14 +391,15 @@ def forward(
         slots: torch.Tensor,
         input_lengths: torch.Tensor,
         max_s: int,
+        true_max_s: int,
         prefill_cache_indices: Optional[torch.Tensor],
     ) -> torch.Tensor:
         hidden_states = self.embed_tokens(input_ids)
 
         # Get rotary cos and sin for this forward
         # Avoid to index in each layer
         cos, sin = self.layers[0].self_attn.rotary_emb.get_cos_sin(
-            position_ids, max_s, hidden_states.dtype
+            position_ids, true_max_s, hidden_states.dtype
         )
 
         residual = None
@@ -449,6 +450,7 @@ def forward(
         prefill_cache_indices: Optional[torch.Tensor],
         lm_head_indices: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
+        true_max_s = max_s
         if prefill_cache_indices is not None:
             # Slots also need to be sliced as it has the same size as the whole kv tensor
             slots = slots[prefill_cache_indices]
@@ -467,6 +469,7 @@ def forward(
             slots,
             input_lengths,
             max_s,
+            true_max_s,
             prefill_cache_indices,
         )
         if lm_head_indices is not None:

server/text_generation_server/models/custom_modeling/flash_mixtral_modeling.py

@@ -401,7 +401,7 @@ def topology(self, x: torch.Tensor, padded_bins: torch.Tensor):
             self.offsets_block_rows = block_rows
             offsets = self.offsets
         else:
-            offsets = self.offsets[:block_rows]
+            offsets = self.offsets[: block_rows + 1]
 
         # Indices for the sparse matrix. The indices for
         # the intermediate matrix are dynamic depending
@@ -632,14 +632,15 @@ def forward(
         slots: torch.Tensor,
         input_lengths: torch.Tensor,
         max_s: int,
+        true_max_s: int,
         prefill_cache_indices: Optional[torch.Tensor],
     ) -> torch.Tensor:
         hidden_states = self.embed_tokens(input_ids)
 
         # Get rotary cos and sin for this forward
         # Avoid to index in each layer
         cos, sin = self.layers[0].self_attn.rotary_emb.get_cos_sin(
-            position_ids, max_s, hidden_states.dtype
+            position_ids, true_max_s, hidden_states.dtype
         )
 
         residual = None
@@ -690,6 +691,7 @@ def forward(
         prefill_cache_indices: Optional[torch.Tensor],
         lm_head_indices: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
+        true_max_s = max_s
         if prefill_cache_indices is not None:
             # Slots also need to be sliced as it has the same size as the whole kv tensor
             slots = slots[prefill_cache_indices]
@@ -708,6 +710,7 @@ def forward(
             slots,
             input_lengths,
             max_s,
+            true_max_s,
             prefill_cache_indices,
         )
         if lm_head_indices is not None: