feat: adjust rotary embed and avoid cuda graphs of size 2 and smaller

integration-tests/models/__snapshots__/test_flash_qwen2_vl_warmup/test_flash_qwen2_vl_simple.json

@@ -0,0 +1,26 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null,
+      "message": {
+        "content": "The correct answer is: blue",
+        "name": null,
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "usage": null
+    }
+  ],
+  "created": 1733445131,
+  "id": "",
+  "model": "Qwen/Qwen2-VL-2B-Instruct",
+  "object": "chat.completion",
+  "system_fingerprint": "2.4.2-dev0-native",
+  "usage": {
+    "completion_tokens": 7,
+    "prompt_tokens": 27,
+    "total_tokens": 34
+  }
+}

integration-tests/models/test_flash_qwen2_vl_warmup.py

@@ -5,7 +5,7 @@
 def flash_qwen2_vl_handle(launcher):
     with launcher(
         "Qwen/Qwen2-VL-2B-Instruct",
-        max_input_tokens=40,
+        max_input_length=40,
         max_batch_prefill_tokens=50,
         max_total_tokens=51,
     ) as handle:

server/text_generation_server/models/__init__.py

@@ -29,6 +29,7 @@
     BloomForCausalLM,
 )
 from text_generation_server.models.globals import ATTENTION
+import text_generation_server.models.globals as globals
 from text_generation_server.models.seq2seq_lm import Seq2SeqLM
 from text_generation_server.models.galactica import GalacticaCausalLMBatch
 from text_generation_server.models.custom_modeling.neox_modeling import (
@@ -1208,6 +1209,11 @@ def get_model(
         else:
             raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
     if model_type == QWEN2_VL:
+        # TODO: remove edge case when cuda graph issue is resolved for BS=2 with Qwen2-VL
+        logger.warning(
+            "Qwen2-VL requires cuda graphs to be greater than 2. Removing all cuda graphs with a batch size equal or less than 2."
+        )
+        globals.CUDA_GRAPHS = list(filter(lambda x: x > 2, globals.CUDA_GRAPHS))
         return VlmCausalLM(
             model_id=model_id,
             model_class=Qwen2VLForConditionalGeneration,

server/text_generation_server/models/custom_modeling/flash_qwen2_modeling.py

@@ -138,7 +138,12 @@ def forward(
                 dim=-1,
             )
 
-        self.rotary_emb(query, torch.select(kv, dim=1, index=0), cos, sin)
+        self.rotary_emb(
+            query,
+            torch.select(kv, dim=1, index=0),
+            cos[: query.shape[0], ...],
+            sin[: query.shape[0], ...],
+        )
 
         if prefill_cache_indices is not None:
             kv_to_cache = kv[prefill_cache_indices]

server/text_generation_server/models/flash_causal_lm.py

@@ -56,11 +56,11 @@
     MEM_POOL,
     ATTENTION,
     BLOCK_SIZE,
-    CUDA_GRAPHS,
     REQUEST_LOGPROBS,
     TGI_WIGGLE_ROOM,
     get_adapter_to_index,
 )
+import text_generation_server.models.globals as globals
 from text_generation_server.layers.attention import KVCache, Seqlen
 from text_generation_server.utils import StoppingCriteria, HeterogeneousNextTokenChooser
 from text_generation_server.utils.dist import MEMORY_FRACTION
@@ -1629,8 +1629,8 @@ def warmup(
                         int(val)
                         for val in os.environ["PYTORCH_TUNABLEOP_SEQLENS"].split(",")
                     ]
-                elif CUDA_GRAPHS is not None:
-                    tuning_sequences = CUDA_GRAPHS
+                elif globals.CUDA_GRAPHS is not None:
+                    tuning_sequences = globals.CUDA_GRAPHS
                 else:
                     tuning_sequences = [1, 2, 3, 4, 5, 6, 7]
 
@@ -1669,20 +1669,22 @@ def warmup(
                     "PyTorch ROCm TunableOp (https://github.com/pytorch/pytorch/tree/main/aten/src/ATen/cuda/tunable) is disabled. TunableOp brings an additional 5-8% latency improvement for small sequence lengths but requires a warmup. If necessary, please use the environment variable PYTORCH_TUNABLEOP_ENABLED=1 to enable TunableOp.",
                 )
 
-        if CUDA_GRAPHS:
+        if globals.CUDA_GRAPHS:
             try:
                 log_master(
-                    logger.info, f"Cuda Graphs are enabled for sizes {CUDA_GRAPHS}"
+                    logger.info,
+                    f"Cuda Graphs are enabled for sizes {globals.CUDA_GRAPHS}",
                 )
                 # Warmup cuda graphs
-                for bs in CUDA_GRAPHS:
+                for bs in globals.CUDA_GRAPHS:
                     if self.speculate is None or self.speculate + 1 <= bs:
                         self.cuda_graph_warmup(bs, max_total_tokens, max_total_tokens)
             except torch.cuda.OutOfMemoryError:
                 logger.exception("Decode cuda graph warmup failed")
         else:
             log_master(
-                logger.info, f"Cuda Graphs are disabled (CUDA_GRAPHS={CUDA_GRAPHS})."
+                logger.info,
+                f"Cuda Graphs are disabled (CUDA_GRAPHS={globals.CUDA_GRAPHS}).",
             )
 
         assert max_input_tokens is not None