feat: improve qwen2-vl startup (#2802)

* feat: tokenize each request individually and increase warmup image size

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

* fix: address image resize and rebase changes

* feat: update to run qwen2-vl tests

* fix: tweak param types

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.py

@@ -1,81 +1,80 @@
-# Disabled because it's broken.
+import pytest
-# import pytest
+
-#
+
-#
+@pytest.fixture(scope="module")
-# @pytest.fixture(scope="module")
+def flash_qwen2_vl_handle(launcher):
-# def flash_qwen2_vl_handle(launcher):
+    with launcher("Qwen/Qwen2-VL-7B-Instruct") as handle:
-#     with launcher("Qwen/Qwen2-VL-7B-Instruct") as handle:
+        yield handle
-#         yield handle
+
-#
+
-#
+@pytest.fixture(scope="module")
-# @pytest.fixture(scope="module")
+async def flash_qwen2(flash_qwen2_vl_handle):
-# async def flash_qwen2(flash_qwen2_vl_handle):
+    await flash_qwen2_vl_handle.health(300)
-#     await flash_qwen2_vl_handle.health(300)
+    return flash_qwen2_vl_handle.client
-#     return flash_qwen2_vl_handle.client
+
-#
+
-#
+@pytest.mark.private
-# @pytest.mark.private
+async def test_flash_qwen2_vl_simple(flash_qwen2, response_snapshot):
-# async def test_flash_qwen2_vl_simple(flash_qwen2, response_snapshot):
+    response = await flash_qwen2.chat(
-#     response = await flash_qwen2.chat(
+        max_tokens=100,
-#         max_tokens=100,
+        seed=42,
-#         seed=42,
+        messages=[
-#         messages=[
+            {
-#             {
+                "role": "user",
-#                 "role": "user",
+                "content": [
-#                 "content": [
+                    {
-#                     {
+                        "type": "image_url",
-#                         "type": "image_url",
+                        "image_url": {
-#                         "image_url": {
+                            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"
-#                             "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"
+                        },
-#                         },
+                    },
-#                     },
+                    {"type": "text", "text": "Describe this image."},
-#                     {"type": "text", "text": "Describe this image."},
+                ],
-#                 ],
+            },
-#             },
+        ],
-#         ],
+    )
-#     )
+
-#
+    assert (
-#     assert (
+        response.choices[0].message.content
-#         response.choices[0].message.content
+        == "The image depicts an anthropomorphic rabbit, wearing a futuristic spacesuit, in an extraterrestrial environment. The setting appears to be a red planet resembling Mars, with rugged terrain and rocky formations in the background. The moon is visible in the distant sky, adding to the lunar landscape."
-#         == "The image depicts an anthropomorphic rabbit, wearing a futuristic spacesuit, in an extraterrestrial environment. The setting appears to be a red planet resembling Mars, with rugged terrain and rocky formations in the background. The moon is visible in the distant sky, adding to the lunar landscape."
+    )
-#     )
+
-#
+    assert response == response_snapshot
-#     assert response == response_snapshot
+
-#
+
-#
+@pytest.mark.private
-# @pytest.mark.private
+async def test_flash_qwen2_vl_simple_streaming(flash_qwen2, response_snapshot):
-# async def test_flash_qwen2_vl_simple_streaming(flash_qwen2, response_snapshot):
+    responses = await flash_qwen2.chat(
-#     responses = await flash_qwen2.chat(
+        max_tokens=100,
-#         max_tokens=100,
+        seed=42,
-#         seed=42,
+        messages=[
-#         messages=[
+            {
-#             {
+                "role": "user",
-#                 "role": "user",
+                "content": [
-#                 "content": [
+                    {
-#                     {
+                        "type": "image_url",
-#                         "type": "image_url",
+                        "image_url": {
-#                         "image_url": {
+                            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"
-#                             "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"
+                        },
-#                         },
+                    },
-#                     },
+                    {"type": "text", "text": "Describe this image."},
-#                     {"type": "text", "text": "Describe this image."},
+                ],
-#                 ],
+            },
-#             },
+        ],
-#         ],
+        stream=True,
-#         stream=True,
+    )
-#     )
+
-#
+    count = 0
-#     count = 0
+    generated = ""
-#     generated = ""
+    last_response = None
-#     last_response = None
+    async for response in responses:
-#     async for response in responses:
+        count += 1
-#         count += 1
+        generated += response.choices[0].delta.content
-#         generated += response.choices[0].delta.content
+        last_response = response
-#         last_response = response
+
-#
+    assert (
-#     assert (
+        generated
-#         generated
+        == "The image depicts an anthropomorphic rabbit, wearing a futuristic spacesuit, in an extraterrestrial environment. The setting appears to be a red planet resembling Mars, with rugged terrain and rocky formations in the background. The moon is visible in the distant sky, adding to the lunar landscape."
-#         == "The image depicts an anthropomorphic rabbit, wearing a futuristic spacesuit, in an extraterrestrial environment. The setting appears to be a red planet resembling Mars, with rugged terrain and rocky formations in the background. The moon is visible in the distant sky, adding to the lunar landscape."
+    )
-#     )
+    assert count == 58
-#     assert count == 58
+    assert last_response == response_snapshot
-#     assert last_response == response_snapshot

integration-tests/models/test_flash_qwen2_vl_warmup.py

@@ -0,0 +1,38 @@
+import pytest
+
+
+@pytest.fixture(scope="module")
+def flash_qwen2_vl_handle(launcher):
+    with launcher(
+        "Qwen/Qwen2-VL-2B-Instruct",
+        max_input_length=40,
+        max_batch_prefill_tokens=50,
+        max_total_tokens=51,
+    ) as handle:
+        yield handle
+
+
+@pytest.fixture(scope="module")
+async def flash_qwen2(flash_qwen2_vl_handle):
+    await flash_qwen2_vl_handle.health(300)
+    return flash_qwen2_vl_handle.client
+
+
+@pytest.mark.private
+async def test_flash_qwen2_vl_simple(flash_qwen2, response_snapshot):
+    response = await flash_qwen2.chat(
+        max_tokens=20,
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is the color of the sky?"},
+                ],
+            },
+        ],
+    )
+
+    assert response.choices[0].message.content == "The correct answer is: blue"
+
+    assert response == response_snapshot

server/text_generation_server/models/__init__.py

@@ -29,6 +29,7 @@ from text_generation_server.models.custom_modeling.bloom_modeling import (
     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 (
@@ -1217,6 +1218,11 @@ 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 @@ class Qwen2Attention(torch.nn.Module):
                 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/custom_modeling/qwen2_vl.py

@@ -517,11 +517,11 @@ class Qwen2VLForConditionalGeneration(nn.Module):
         pixel_values: torch.FloatTensor = None,
         image_grid_thw: Optional[torch.LongTensor] = None,
         video_grid_thw: Optional[torch.LongTensor] = None,
-        pixel_attention_mask=None,
+        pixel_attention_mask: Optional[torch.Tensor] = None,
         image_sizes: Optional[torch.LongTensor] = None,
         adapter_data: Optional[torch.Tensor] = None,
         cross_attention_states: Optional[torch.Tensor] = None,
-        image_indices=None,
+        image_indices: Optional[torch.Tensor] = None,
     ):
         inputs_embeds = self.embed_tokens(input_ids)
 
@@ -533,6 +533,7 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 ).squeeze(0)
                 inputs_embeds[input_ids == self.image_token_id] = image_embeds
 
+        max_s = max(max_s, inputs_embeds.size(0))
         hidden_states = self.text_model(
             inputs_embeds=inputs_embeds,
             position_ids=position_ids,

server/text_generation_server/models/flash_causal_lm.py

@@ -56,11 +56,13 @@ from text_generation_server.models.globals import (
     MEM_POOL,
     ATTENTION,
     BLOCK_SIZE,
-    CUDA_GRAPHS,
     REQUEST_LOGPROBS,
     TGI_WIGGLE_ROOM,
     get_adapter_to_index,
 )
+
+# avoid coping CUDA_GRAPHS value by importing globals as a module
+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
@@ -1635,8 +1637,8 @@ class FlashCausalLM(Model):
                         int(val)
                         for val in os.environ["PYTORCH_TUNABLEOP_SEQLENS"].split(",")
                     ]
-                elif CUDA_GRAPHS is not None:
+                elif globals.CUDA_GRAPHS is not None:
-                    tuning_sequences = CUDA_GRAPHS
+                    tuning_sequences = globals.CUDA_GRAPHS
                 else:
                     tuning_sequences = [1, 2, 3, 4, 5, 6, 7]
 
@@ -1675,13 +1677,14 @@ class FlashCausalLM(Model):
                     "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:
                     synchronize(self.device)
                     free_memory = get_free_memory(
                         self.device, MEMORY_FRACTION * TGI_WIGGLE_ROOM
@@ -1705,7 +1708,8 @@ class FlashCausalLM(Model):
                 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

server/text_generation_server/models/vlm_causal_lm.py

@@ -236,7 +236,6 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
                             w = image.width * 2
                             h = image.height * 2
                             image = image.resize((w, h))
-
                     if config.model_type == "llava_next":
                         images.append(image)
                     else: