fix cuda graphs for qwen2-vl (#2708)

* feat: support multidimensional position ids on batch to enable cuda graphs on qwen2-vl

* fix: only check model type if config exists

* fix: adjust sharding and lm head logic

* fix qwen2 failure in intel cpu

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* fix: return correct shape logits and add streaming test

* fix: remove unused import and refactor test

---------

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

integration-tests/models/__snapshots__/test_flash_qwen2_vl/test_flash_qwen2_vl_simple_streaming.json

@@ -0,0 +1,20 @@
+{
+  "choices": [
+    {
+      "delta": {
+        "content": "",
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null
+    }
+  ],
+  "created": 1730416361,
+  "id": "",
+  "model": "Qwen/Qwen2-VL-7B-Instruct",
+  "object": "chat.completion.chunk",
+  "system_fingerprint": "2.4.1-dev0-native",
+  "usage": null
+}

integration-tests/models/test_flash_qwen2_vl.py

@@ -3,7 +3,7 @@ import pytest
 
 @pytest.fixture(scope="module")
 def flash_qwen2_vl_handle(launcher):
-    with launcher("Qwen/Qwen2-VL-7B-Instruct", cuda_graphs=[0]) as handle:
+    with launcher("Qwen/Qwen2-VL-7B-Instruct") as handle:
         yield handle
 
 
@@ -40,3 +40,41 @@ async def test_flash_qwen2_vl_simple(flash_qwen2, response_snapshot):
     )
 
     assert response == response_snapshot
+
+
+@pytest.mark.private
+async def test_flash_qwen2_vl_simple_streaming(flash_qwen2, response_snapshot):
+    responses = await flash_qwen2.chat(
+        max_tokens=100,
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image_url",
+                        "image_url": {
+                            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"
+                        },
+                    },
+                    {"type": "text", "text": "Describe this image."},
+                ],
+            },
+        ],
+        stream=True,
+    )
+
+    count = 0
+    generated = ""
+    last_response = None
+    async for response in responses:
+        count += 1
+        generated += response.choices[0].delta.content
+        last_response = response
+
+    assert (
+        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."
+    )
+    assert count == 58
+    assert last_response == response_snapshot

server/text_generation_server/models/custom_modeling/qwen2_vl.py

@@ -34,7 +34,7 @@ from text_generation_server.layers import (
     TensorParallelColumnLinear,
     TensorParallelRowLinear,
     TensorParallelEmbedding,
-    FastLinear,
+    SpeculativeHead,
 )
 from text_generation_server.layers.attention import (
     Seqlen,
@@ -69,7 +69,7 @@ def apply_rotary_pos_emb_vision(
 class Qwen2VLSdpaAttention(nn.Module):
     def __init__(self, *, prefix, config, weights):
         super().__init__()
-        self.embed_dim = config.embed_dim
+        self.embed_dim = config.embed_dim // weights.process_group.size()
         self.head_dim = config.hidden_size // config.num_heads
         self.num_heads = config.num_heads // weights.process_group.size()
 
@@ -82,7 +82,7 @@ class Qwen2VLSdpaAttention(nn.Module):
             num_key_value_heads=self.num_heads,
         )
         self.qkv.linear.bias = weights.get_sharded(f"{prefix}.qkv.bias", dim=0)
-        self.proj = TensorParallelColumnLinear.load(
+        self.proj = TensorParallelRowLinear.load(
             config,
             prefix=f"{prefix}.proj",
             weights=weights,
@@ -364,8 +364,15 @@ class Qwen2VLForConditionalGeneration(nn.Module):
             prefix="visual", config=config.vision_config, weights=weights
         )
         self.text_model = Qwen2Model(prefix=None, config=config, weights=weights)
-        self.lm_head = FastLinear.load(
+        if config.tie_word_embeddings:
-            prefix="lm_head", weights=weights, config=config, bias=False
+            suffix = "model.embed_tokens"
+        else:
+            suffix = "lm_head"
+
+        self.lm_head = SpeculativeHead.load(
+            config,
+            prefix=suffix if not prefix else f"{prefix}.{suffix}",
+            weights=weights,
         )
         self.norm = FastRMSNorm.load(
             prefix="model.norm",
@@ -377,9 +384,12 @@ class Qwen2VLForConditionalGeneration(nn.Module):
     def get_position_ids(
         self,
         batch_input_ids: torch.Tensor,
-        image_grid_thw: Optional[torch.LongTensor],
+        image_grid_thw: Optional[torch.LongTensor] = None,
         # video_grid_thw is not implemented yet as we do not accept video inputs at the moment
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if batch_input_ids.dim() == 1:
+            batch_input_ids = batch_input_ids.unsqueeze(0)
+
         position_ids = torch.ones(
             3,
             batch_input_ids.shape[0],
@@ -505,5 +515,7 @@ class Qwen2VLForConditionalGeneration(nn.Module):
             prefill_cache_indices=prefill_cache_indices,
         )
         hidden_states, _ = self.norm(hidden_states)
-        logits = self.lm_head(hidden_states)
+        if lm_head_indices is not None:
-        return logits, None
+            hidden_states = hidden_states[lm_head_indices]
+        logits, speculative_logits = self.lm_head(hidden_states)
+        return logits, speculative_logits

server/text_generation_server/models/flash_causal_lm.py

@@ -1430,6 +1430,14 @@ class FlashCausalLM(Model):
         else:
             state = None
 
+        if (
+            hasattr(self.model, "config")
+            and hasattr(self.model.config, "model_type")
+            and self.model.config.model_type == "qwen2_vl"
+        ):
+            if position_ids.dim() == 1:
+                position_ids = self.model.get_position_ids(input_ids)
+
         graph = torch.cuda.CUDAGraph()
         self.cuda_graphs[bs] = {
             "input_ids": input_ids,
@@ -1806,7 +1814,7 @@ class FlashCausalLM(Model):
         # Copy inputs to the static inputs of the cuda graph
         # Static inputs are potentially padded
         cuda_graph["input_ids"][: input_ids.shape[0]] = input_ids
-        cuda_graph["position_ids"][: position_ids.shape[0]] = position_ids
+        cuda_graph["position_ids"][: position_ids.shape[-1]] = position_ids
         if ATTENTION == "flashinfer":
             block_tables = block_tables_to_ragged(
                 block_tables=block_tables,
@@ -1981,7 +1989,7 @@ class FlashCausalLM(Model):
         # instantly become of shape [BATCH_SIZE]
         if prefill and finished_prefilling:
             indices = batch.cu_seqlen_prefill[1:] - 1
-            batch.position_ids = batch.position_ids[indices]
+            batch.position_ids = batch.position_ids[(..., indices)]
             batch.slot_indices = batch.slot_indices[indices]
             batch.adapter_meta.adapter_indices = batch.adapter_meta.adapter_indices[
                 indices

server/text_generation_server/models/vlm_causal_lm.py

@@ -363,15 +363,12 @@ class VlmCausalLM(FlashCausalLM):
             max_s = batch.max_current_length
             lm_head_indices = batch.prefill_head_indices
 
-        if hasattr(self.model, "get_position_ids"):
+        if self.model.config.model_type == "qwen2_vl":
-            if position_ids.shape[0] != 1:
+            if position_ids.dim() == 1 and batch.prefilling:
                 position_ids = self.model.get_position_ids(
-                    input_ids.unsqueeze(0), batch.image_grid_thw
+                    input_ids, batch.image_grid_thw
                 )
-                batch.position_ids = position_ids[0, 0, :]
+                batch.position_ids = position_ids
-            else:
-                position_ids = position_ids.repeat(3, 1, 1).clone()
-                batch.position_ids = position_ids[0, 0, :]
 
         if cu_seqlen_prefill is None and self.max_past() is not None:
             # In decode, not prefill, we're actually overwriting the KV-cache