[Kernel][Model] logits_soft_cap for Gemma2 with flashinfer (#6051)

Co-authored-by: Simon Mo <simon.mo@hey.com>

.buildkite/test-pipeline.yaml

@@ -118,12 +118,15 @@ steps:
 
 - label: Kernels Test %N
   #mirror_hardwares: [amd]
-  command: pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  commands:
+    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.7/flashinfer-0.0.7+cu121torch2.3-cp310-cp310-linux_x86_64.whl
+    - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 4
 
 - label: Models Test
   #mirror_hardwares: [amd]
   commands:
+    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.7/flashinfer-0.0.7+cu121torch2.3-cp310-cp310-linux_x86_64.whl
     - pytest -v -s models -m \"not vlm\"
 
 - label: Vision Language Models Test
@@ -234,7 +237,7 @@ steps:
   - pytest -v -s distributed/test_custom_all_reduce.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
-  - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.5/flashinfer-0.0.5+cu121torch2.3-cp310-cp310-linux_x86_64.whl
+  - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.7/flashinfer-0.0.7+cu121torch2.3-cp310-cp310-linux_x86_64.whl
   - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=meta-llama/Meta-Llama-3-8B DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - pytest -v -s -x lora/test_mixtral.py

tests/kernels/test_flashinfer.py

+from typing import List, Optional, Tuple
+
+import flashinfer
+import pytest
+import torch
+
+NUM_HEADS = [(16, 16), (32, 8), (64, 8)]
+HEAD_SIZES = [128, 256]
+BLOCK_SIZES = [16, 32]
+DTYPES = [torch.float16, torch.bfloat16]
+NUM_BLOCKS = 32768  # Large enough to test overflow in index calculation.
+
+
+def ref_paged_attn(
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    query_lens: List[int],
+    kv_lens: List[int],
+    block_tables: torch.Tensor,
+    scale: float,
+    sliding_window: Optional[int] = None,
+    soft_cap: Optional[float] = None,
+) -> torch.Tensor:
+    num_seqs = len(query_lens)
+    block_tables = block_tables.cpu().numpy()
+    _, block_size, num_kv_heads, head_size = key_cache.shape
+
+    outputs: List[torch.Tensor] = []
+    start_idx = 0
+    for i in range(num_seqs):
+        query_len = query_lens[i]
+        kv_len = kv_lens[i]
+        q = query[start_idx:start_idx + query_len]
+        q *= scale
+
+        num_kv_blocks = (kv_len + block_size - 1) // block_size
+        block_indices = block_tables[i, :num_kv_blocks]
+
+        k = key_cache[block_indices].view(-1, num_kv_heads, head_size)
+        k = k[:kv_len]
+        v = value_cache[block_indices].view(-1, num_kv_heads, head_size)
+        v = v[:kv_len]
+
+        if q.shape[1] != k.shape[1]:
+            k = torch.repeat_interleave(k, q.shape[1] // k.shape[1], dim=1)
+            v = torch.repeat_interleave(v, q.shape[1] // v.shape[1], dim=1)
+        attn = torch.einsum("qhd,khd->hqk", q, k).float()
+        empty_mask = torch.ones(query_len, kv_len)
+        mask = torch.triu(empty_mask, diagonal=kv_len - query_len + 1).bool()
+        if sliding_window is not None:
+            sliding_window_mask = torch.triu(empty_mask,
+                                             diagonal=kv_len -
+                                             (query_len + sliding_window) +
+                                             1).bool().logical_not()
+            mask |= sliding_window_mask
+        if soft_cap is not None:
+            attn = soft_cap * torch.tanh(attn / soft_cap)
+        attn.masked_fill_(mask, float("-inf"))
+        attn = torch.softmax(attn, dim=-1).to(v.dtype)
+        out = torch.einsum("hqk,khd->qhd", attn, v)
+
+        outputs.append(out)
+        start_idx += query_len
+
+    return torch.cat(outputs, dim=0)
+
+
+@pytest.mark.parametrize("kv_lens", [[1328, 18, 463], [1, 54, 293, 70]])
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("soft_cap", [None, 30.0, 50.0])
+@torch.inference_mode
+def test_flashinfer_decode_with_paged_kv(kv_lens: List[int],
+                                         num_heads: Tuple[int,
+                                                          int], head_size: int,
+                                         dtype: torch.dtype, block_size: int,
+                                         soft_cap: Optional[float]) -> None:
+    torch.set_default_device("cuda")
+    torch.cuda.manual_seed_all(0)
+    num_seqs = len(kv_lens)
+    num_query_heads = num_heads[0]
+    num_kv_heads = num_heads[1]
+    assert num_query_heads % num_kv_heads == 0
+    max_kv_len = max(kv_lens)
+    scale = head_size**-0.5
+
+    query = torch.randn(num_seqs, num_query_heads, head_size, dtype=dtype)
+    key_value_cache = torch.randn(NUM_BLOCKS,
+                                  2,
+                                  block_size,
+                                  num_kv_heads,
+                                  head_size,
+                                  dtype=dtype)
+    key_cache = key_value_cache[:, 0, :, :, :].squeeze(1)
+    value_cache = key_value_cache[:, 1, :, :, :].squeeze(1)
+
+    max_num_blocks_per_seq = (max_kv_len + block_size - 1) // block_size
+    block_tables = torch.randint(0,
+                                 NUM_BLOCKS,
+                                 (num_seqs, max_num_blocks_per_seq),
+                                 dtype=torch.int32)
+
+    kv_indptr = [0]
+    kv_indices = []
+    kv_last_page_lens = []
+    for i in range(num_seqs):
+        seq_len = kv_lens[i]
+        assert seq_len > 0
+        num_blocks = (seq_len + block_size - 1) // block_size
+        kv_indices.extend(block_tables[i, :num_blocks])
+        kv_indptr.append(kv_indptr[-1] + num_blocks)
+        kv_last_page_len = seq_len % block_size
+        if kv_last_page_len == 0:
+            kv_last_page_len = block_size
+        kv_last_page_lens.append(kv_last_page_len)
+
+    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
+    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
+    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
+
+    workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8)
+    wrapper = flashinfer.\
+        BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, "NHD")
+    wrapper.begin_forward(kv_indptr,
+                          kv_indices,
+                          kv_last_page_lens,
+                          num_query_heads,
+                          num_kv_heads,
+                          head_size,
+                          block_size,
+                          "NONE",
+                          data_type=dtype)
+
+    output = wrapper.forward(query, key_value_cache, logits_soft_cap=soft_cap)
+
+    ref_output = ref_paged_attn(query=query,
+                                key_cache=key_cache,
+                                value_cache=value_cache,
+                                query_lens=[1] * num_seqs,
+                                kv_lens=kv_lens,
+                                block_tables=block_tables,
+                                scale=scale,
+                                soft_cap=soft_cap)
+    assert torch.allclose(output, ref_output, atol=1e-2, rtol=1e-2), \
+        f"{torch.max(torch.abs(output - ref_output))}"
+
+
+@pytest.mark.parametrize("seq_lens", [[(1, 1328), (5, 18), (129, 463)]])
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("soft_cap", [None, 30.0, 50.0])
+@torch.inference_mode
+def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
+                                          num_heads: Tuple[int, int],
+                                          head_size: int, dtype: torch.dtype,
+                                          block_size: int,
+                                          soft_cap: Optional[float]) -> None:
+    torch.set_default_device("cuda")
+    torch.cuda.manual_seed_all(0)
+    num_seqs = len(seq_lens)
+    query_lens = [x[0] for x in seq_lens]
+    kv_lens = [x[1] for x in seq_lens]
+    num_query_heads = num_heads[0]
+    num_kv_heads = num_heads[1]
+    assert num_query_heads % num_kv_heads == 0
+    max_kv_len = max(kv_lens)
+    scale = head_size**-0.5
+
+    query = torch.randn(sum(query_lens),
+                        num_query_heads,
+                        head_size,
+                        dtype=dtype)
+    key_value_cache = torch.randn(NUM_BLOCKS,
+                                  2,
+                                  block_size,
+                                  num_kv_heads,
+                                  head_size,
+                                  dtype=dtype)
+    key_cache = key_value_cache[:, 0, :, :, :].squeeze(1)
+    value_cache = key_value_cache[:, 1, :, :, :].squeeze(1)
+
+    # Normalize the scale of the key and value caches to mitigate
+    # numerical instability.
+    key_cache /= head_size**0.5
+    value_cache /= head_size**0.5
+
+    max_num_blocks_per_seq = (max_kv_len + block_size - 1) // block_size
+    block_tables = torch.randint(0,
+                                 NUM_BLOCKS,
+                                 (num_seqs, max_num_blocks_per_seq),
+                                 dtype=torch.int32)
+
+    qo_indptr = [0]
+    kv_indptr = [0]
+    kv_indices = []
+    kv_last_page_lens = []
+    for i in range(num_seqs):
+        seq_len = kv_lens[i]
+        assert seq_len > 0
+        num_blocks = (seq_len + block_size - 1) // block_size
+        kv_indices.extend(block_tables[i, :num_blocks])
+        kv_indptr.append(kv_indptr[-1] + num_blocks)
+        kv_last_page_len = seq_len % block_size
+        if kv_last_page_len == 0:
+            kv_last_page_len = block_size
+        kv_last_page_lens.append(kv_last_page_len)
+        qo_indptr.append(qo_indptr[-1] + query_lens[i])
+
+    qo_indptr = torch.tensor(qo_indptr, dtype=torch.int32)
+    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
+    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
+    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
+
+    workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8)
+    wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
+        workspace_buffer, "NHD")
+    wrapper.begin_forward(
+        qo_indptr,
+        kv_indptr,
+        kv_indices,
+        kv_last_page_lens,
+        num_query_heads,
+        num_kv_heads,
+        head_size,
+        block_size,
+    )
+
+    output = wrapper.forward(
+        query,
+        key_value_cache,
+        logits_soft_cap=soft_cap,
+    )
+
+    ref_output = ref_paged_attn(query=query,
+                                key_cache=key_cache,
+                                value_cache=value_cache,
+                                query_lens=query_lens,
+                                kv_lens=kv_lens,
+                                block_tables=block_tables,
+                                scale=scale,
+                                soft_cap=soft_cap)
+    assert torch.allclose(output, ref_output, atol=1e-2, rtol=1e-2), \
+        f"{torch.max(torch.abs(output - ref_output))}"

vllm/attention/backends/flashinfer.py

@@ -102,6 +102,8 @@ class FlashInferMetadata(AttentionMetadata):
     # The data type of the paged kv cache
     data_type: torch.dtype = None
     device: torch.device = torch.device("cuda")
+    # Only used by gemma2 model
+    logits_soft_cap: Optional[float] = None
 
     def __post_init__(self):
         # Refer to
@@ -271,9 +273,11 @@ class FlashInferImpl(AttentionImpl):
             else:
                 assert prefill_meta is not None
                 assert prefill_meta.prefill_wrapper is not None
-                output = prefill_meta.prefill_wrapper.forward(query,
-                                                              kv_cache,
-                                                              causal=True)
+                output = prefill_meta.prefill_wrapper.forward(
+                    query,
+                    kv_cache,
+                    logits_soft_cap=attn_metadata.logits_soft_cap,
+                    causal=True)
         else:
             assert attn_metadata.decode_metadata is not None
             assert attn_metadata.decode_metadata.decode_wrapper is not None
@@ -281,5 +285,5 @@ class FlashInferImpl(AttentionImpl):
                 query,
                 kv_cache,
                 sm_scale=self.scale,
-            )
+                logits_soft_cap=attn_metadata.logits_soft_cap)
         return output.view(num_tokens, hidden_size)

vllm/attention/selector.py

@@ -77,9 +77,9 @@ def get_attn_backend(
         return IpexAttnBackend
     elif backend == _Backend.FLASHINFER:
         logger.info("Using Flashinfer backend.")
-        logger.warning(("Flashinfer will be stuck on llma-2-7b,"
-                        " please avoid using Flashinfer as the"
-                        "backend when running on llma-2-7b."))
+        logger.warning(("Flashinfer will be stuck on llama-2-7b,"
+                        " please avoid using Flashinfer as the "
+                        "backend when running on llama-2-7b."))
         from vllm.attention.backends.flashinfer import FlashInferBackend
         return FlashInferBackend
     elif backend == _Backend.PALLAS:

vllm/model_executor/models/gemma2.py

@@ -38,7 +38,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors, SamplerOutput
-from vllm.utils import print_warning_once
 
 from .interfaces import SupportsLoRA
 
@@ -137,12 +136,6 @@ class Gemma2Attention(nn.Module):
             dtype=torch.get_default_dtype(),
         )
 
-        if self.config.attn_logit_softcapping is not None:
-            print_warning_once(
-                "Gemma 2 normally uses attention logit soft-capping; "
-                "soft-capping is currently incompatible with the flash "
-                "attention kernels, so vLLM removes it to enable speed and "
-                "efficiency gains of flash attention.")
         # FIXME(woosuk): While Gemma 2 uses sliding window attention for every
         # odd layer, vLLM currently ignores it and uses global attention for
         # all layers.

vllm/worker/model_runner.py

@@ -15,7 +15,7 @@ try:
     from flashinfer import BatchDecodeWithPagedKVCacheWrapper
     from flashinfer.decode import CUDAGraphBatchDecodeWithPagedKVCacheWrapper
     from flashinfer.prefill import BatchPrefillWithPagedKVCacheWrapper
-    FLASHINFER_WORKSPACE_BUFFER_SIZE = 128 * 1024 * 1024
+    FLASHINFER_WORKSPACE_BUFFER_SIZE = 256 * 1024 * 1024
 except ImportError:
     BatchDecodeWithPagedKVCacheWrapper = None
     CUDAGraphBatchDecodeWithPagedKVCacheWrapper = None
@@ -683,6 +683,16 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                                            dtype=torch.long,
                                            device=self.device)
 
+        logits_soft_cap = getattr(self.model_config.hf_config,
+                                  'attn_logit_softcapping', None)
+        if logits_soft_cap is not None and self.attn_backend.get_name(
+        ) != "flashinfer":
+            raise ValueError("Please use Flashinfer backend for models with"
+                             "logits_soft_cap (i.e., Gemma-2)."
+                             " Otherwise, the output might be wrong."
+                             " Set Flashinfer backend by "
+                             "export VLLM_ATTENTION_BACKEND=FLASHINFER.")
+
         if self.attn_backend.get_name() == "flashinfer":
             if len(paged_kv_indptr) > 0:
                 paged_kv_indices_tensor = torch.tensor(paged_kv_indices,
@@ -700,7 +710,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
 
             kv_cache_dtype = get_kv_cache_torch_dtype(self.kv_cache_dtype,
                                                       self.model_config.dtype)
-
             attn_metadata = self.attn_backend.make_metadata(
                 num_prefills=num_prefills,
                 slot_mapping=slot_mapping_tensor,
@@ -721,7 +730,8 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                 query_start_loc=query_start_loc,
                 device=self.device,
                 data_type=kv_cache_dtype,
-                use_cuda_graph=use_captured_graph)
+                use_cuda_graph=use_captured_graph,
+                logits_soft_cap=logits_soft_cap)
 
         else:
             attn_metadata = self.attn_backend.make_metadata(
@@ -1196,7 +1206,8 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
             if model_input.attn_metadata.use_cuda_graph:
                 batch_size = model_input.input_tokens.shape[0]
                 model_input.attn_metadata.decode_wrapper = self.graph_runners[
-                    batch_size].flashinfer_decode_wrapper
+                    model_input.
+                    virtual_engine][batch_size].flashinfer_decode_wrapper
             else:
                 model_input.attn_metadata.decode_wrapper = \
                     self.flashinfer_decode_wrapper