feat: add functional per-head FP8 quantization for FA3 (#1033)

This PR adds FP8 support in FA3 to speed up compute-bound prefill
kernels. It follows up on
#869.

## 1. Bug fixes
- Fixed deadlock, illegal memory access, and wrong results for varied
`num_heads` and `seq_len`.
- Covered by [unit
tests](https://github.com/happierpig/flashinfer-ai/blob/082cf2402fe08332a147f447ef54b75332a20f57/tests/test_hopper_fp8_attention.py).

## 2. New features
- Enabled FP8 in-kernel transpose logic of `mainloop_sparse.cuh`.
- FP8 now works in:
  - `BatchPrefillWithPagedKVCache`
- `BlockSparseAttentionWrapper`: support _**sparse**_ and
_**quantized**_ attention

## 3. Python JIT interface
- Exposed kernels to Python:
  - `BlockSparseAttentionWrapper`
  - `single_prefill_with_kv_cache`
- Migrated tests and benchmarks to Python scripts:
-
[tests/test_hopper_fp8_attention.py](https://github.com/happierpig/flashinfer-ai/blob/082cf2402fe08332a147f447ef54b75332a20f57/tests/test_hopper_fp8_attention.py)
-
[benchmarks/bench_hopper_fp8_attention.py](https://github.com/happierpig/flashinfer-ai/blob/082cf2402fe08332a147f447ef54b75332a20f57/benchmarks/bench_hopper_fp8_attention.py)

**Note:** Performance is on par with
#869. Need tuning.

cc @yzh119

---------

Co-authored-by: happierpig <[email protected]>

Author: happierpig

benchmarks/bench_hopper_fp8_attention.py

@@ -0,0 +1,58 @@
+import torch
+import triton
+
+import flashinfer
+
+
+def bench_single_prefill(seq_len, num_heads, causal, head_dim):
+    num_qo_heads = num_kv_heads = num_heads
+    q = torch.randn(seq_len, num_qo_heads, head_dim, dtype=torch.half, device="cuda")
+    k = torch.randn(seq_len, num_kv_heads, head_dim, dtype=torch.half, device="cuda")
+    v = torch.randn(seq_len, num_kv_heads, head_dim, dtype=torch.half, device="cuda")
+
+    sm80_ms, sm90_ms = (
+        triton.testing.do_bench(
+            lambda: flashinfer.single_prefill_with_kv_cache_return_lse(
+                q, k, v, causal=causal, backend=backend
+            ),
+            warmup=100,
+            rep=1000,
+        )
+        for backend in ["fa2", "fa3"]
+    )
+
+    q = torch.randn(
+        seq_len, num_qo_heads, head_dim, dtype=torch.half, device="cuda"
+    ).to(dtype=torch.float8_e4m3fn)
+    k = torch.randn(
+        seq_len, num_kv_heads, head_dim, dtype=torch.half, device="cuda"
+    ).to(dtype=torch.float8_e4m3fn)
+    v = torch.randn(
+        seq_len, num_kv_heads, head_dim, dtype=torch.half, device="cuda"
+    ).to(dtype=torch.float8_e4m3fn)
+
+    fp8_sm90_ms = triton.testing.do_bench(
+        lambda: flashinfer.single_prefill_with_kv_cache_return_lse(
+            q, k, v, causal=causal, backend="fa3", o_dtype=torch.half
+        ),
+        warmup=100,
+        rep=1000,
+    )
+
+    def flops(ms):
+        if causal:
+            return seq_len * seq_len * num_qo_heads * head_dim * 2 / ms / 1e9
+        else:
+            return seq_len * seq_len * num_qo_heads * head_dim * 4 / ms / 1e9
+
+    print(
+        f"bench_single_prefill (seq_len={seq_len}, num_heads={num_heads}, causal={causal}, head_dim={head_dim}), fa2-template: {flops(sm80_ms):.3f} TFLOPs/s, fa3-template: {flops(sm90_ms):.3f} TFLOPs/s, fa3-fp8: {flops(fp8_sm90_ms):.3f} TFLOPs/s"
+    )
+
+
+if __name__ == "__main__":
+    for seq_len in [4096, 8192, 16384]:
+        for num_heads in [24, 32]:
+            for causal in [True, False]:
+                for head_dim in [64, 128, 256]:
+                    bench_single_prefill(seq_len, num_heads, causal, head_dim)

csrc/batch_prefill_fp8_paged_sm90_kernel_inst.jinja

@@ -0,0 +1,15 @@
+#include <flashinfer/attention/hopper/quantization/prefill_sm90.cuh>
+#include "batch_prefill_sm90_config.inc"
+
+namespace flashinfer {
+
+{% for same_scheduler_for_all_heads in ["true", "false"] %}
+template cudaError_t BatchFP8PrefillWithPagedKVCacheDispatched
+    <{{ head_dim_qk }},
+     {{ mask_mode }},
+     /*USE_SLIDING_WINDOW=*/{{ use_sliding_window }},
+     /*SAME_SCHEDULER_FOR_ALL_HEADS=*/{{ same_scheduler_for_all_heads }},
+     {{ variant_name }}, PagedParams>(PagedParams& params, cudaStream_t stream);
+{% endfor %}
+
+};  // namespace flashinfer

csrc/batch_prefill_fp8_ragged_sm90_kernel_inst.jinja

@@ -0,0 +1 @@
+// TODO: Not implemented yet

csrc/batch_prefill_fp8_sm90.cu

@@ -0,0 +1,179 @@
+/*
+ * Copyright (c) 2023 by FlashInfer team.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <flashinfer/attention/mask.cuh>
+#include <flashinfer/attention/scheduler.cuh>
+#include <flashinfer/layout.cuh>
+#include <flashinfer/math.cuh>
+#include <optional>
+
+#include "batch_prefill_sm90_config.inc"
+#include "pytorch_conversion_utils.h"
+#include "pytorch_extension_utils.h"
+
+namespace flashinfer {
+
+template <uint32_t HEAD_DIM, MaskMode MASK_MODE, bool LEFT_SLIDING_WINDOW,
+          bool SAME_SCHEDULE_FOR_ALL_HEADS, typename AttentionVariant, typename Params>
+cudaError_t BatchFP8PrefillWithPagedKVCacheDispatched(Params& params, cudaStream_t stream);
+
+}  // namespace flashinfer
+
+using namespace flashinfer;
+
+at::Tensor BatchPrefillWithKVCacheSM90Plan(
+    at::Tensor float_workspace_buffer, at::Tensor int_workspace_buffer,
+    at::Tensor page_locked_int_workspace_buffer, at::Tensor qo_indptr, at::Tensor kv_indptr,
+    at::Tensor kv_len_arr, int64_t total_num_rows, int64_t batch_size, int64_t num_qo_heads,
+    int64_t num_kv_heads, int64_t page_size, bool enable_cuda_graph, int64_t head_dim_qk,
+    int64_t head_dim_vo, bool causal) {
+  size_t float_workspace_size_in_bytes =
+      float_workspace_buffer.size(0) * float_workspace_buffer.element_size();
+  size_t int_workspace_size_in_bytes =
+      int_workspace_buffer.size(0) * int_workspace_buffer.element_size();
+
+  flashinfer::PrefillPlanSM90Info plan_info;
+
+  const c10::cuda::OptionalCUDAGuard device_guard(float_workspace_buffer.device());
+  cudaStream_t stream = c10::cuda::getCurrentCUDAStream();
+
+  cudaError_t status =
+      PrefillSM90Plan(float_workspace_buffer.data_ptr(), float_workspace_size_in_bytes,
+                      int_workspace_buffer.data_ptr(), page_locked_int_workspace_buffer.data_ptr(),
+                      int_workspace_size_in_bytes, plan_info, qo_indptr.data_ptr<IdType>(),
+                      kv_indptr.data_ptr<IdType>(), kv_len_arr.data_ptr<IdType>(), total_num_rows,
+                      batch_size, num_qo_heads, num_kv_heads, head_dim_qk, head_dim_vo, page_size,
+                      causal, enable_cuda_graph, /*sizeof_dtype_o=*/2, stream);
+
+  TORCH_CHECK(status == cudaSuccess,
+              "PrefillSM90Plan failed with error: ", cudaGetErrorString(status));
+
+  return vec_to_tensor(plan_info.ToVector());
+}
+
+void BatchPrefillWithRaggedKVCacheSM90Run(at::Tensor float_workspace_buffer,
+                                          at::Tensor int_workspace_buffer, at::Tensor plan_info_vec,
+                                          at::Tensor q, at::Tensor k, at::Tensor v,
+                                          at::Tensor qo_indptr, at::Tensor kv_indptr, at::Tensor o,
+                                          std::optional<at::Tensor> maybe_lse,
+                                          int64_t mask_mode_code, int64_t layout,
+                                          int64_t window_left ADDITIONAL_FUNC_PARAMS) {
+  return;  // TODO: Implement this function
+}
+
+void BatchPrefillWithPagedKVCacheSM90Run(
+    at::Tensor float_workspace_buffer, at::Tensor int_workspace_buffer, at::Tensor plan_info_vec,
+    at::Tensor q, at::Tensor paged_k_cache, at::Tensor paged_v_cache, at::Tensor qo_indptr,
+    at::Tensor paged_kv_indptr, at::Tensor paged_kv_indices, at::Tensor paged_kv_last_page_len,
+    at::Tensor o, std::optional<at::Tensor> maybe_lse, int64_t mask_mode_code, int64_t layout,
+    int64_t window_left ADDITIONAL_FUNC_PARAMS) {
+  PrefillPlanSM90Info plan_info;
+  plan_info.FromVector(tensor_to_vec(plan_info_vec));
+
+  if (maybe_lse) {
+    const auto& lse = *maybe_lse;
+    TORCH_CHECK(lse.size(0) == q.size(0), lse.size(0), q.size(0));
+    TORCH_CHECK(lse.size(1) == q.size(1), lse.size(1), q.size(1));
+  }
+  QKVLayout kv_layout = static_cast<QKVLayout>(layout);
+  int64_t num_kv_heads, page_size;
+  int64_t head_dim_qk = q.size(2);
+  int64_t head_dim_vo = paged_v_cache.size(3);
+  if (kv_layout == QKVLayout::kHND) {
+    num_kv_heads = paged_k_cache.size(1);
+    page_size = paged_k_cache.size(2);
+  } else {
+    page_size = paged_k_cache.size(1);
+    num_kv_heads = paged_k_cache.size(2);
+  }
+
+  void* float_buffer_ptr = float_workspace_buffer.data_ptr();
+  void* int_buffer_ptr = int_workspace_buffer.data_ptr();
+
+  auto q_scalar_type = q.scalar_type();
+  auto kv_scalar_type = paged_k_cache.scalar_type();
+
+  const c10::cuda::OptionalCUDAGuard device_guard(float_workspace_buffer.device());
+  cudaStream_t stream = c10::cuda::getCurrentCUDAStream();
+  const MaskMode mask_mode = static_cast<MaskMode>(mask_mode_code);
+  bool use_swa = window_left != -1;
+
+  DISPATCH_context(
+      DTypeQ, DTypeKV, DTypeO, IdType, MASK_MODE, HEAD_DIM_QK, HEAD_DIM_VO, USE_SLIDING_WINDOW,
+      USE_LOGITS_SOFT_CAP, AttentionVariant, RaggedParams, PagedParams, [&] {
+        PagedParams params;
+
+        params.q_ptr = static_cast<DTypeQ*>(q.data_ptr());
+        params.k_ptr = static_cast<DTypeKV*>(paged_k_cache.data_ptr());
+        params.v_ptr = static_cast<DTypeKV*>(paged_v_cache.data_ptr());
+        params.o_ptr = static_cast<DTypeO*>(o.data_ptr());
+        params.lse_ptr = maybe_lse ? static_cast<float*>(maybe_lse->data_ptr()) : nullptr;
+        params.q_stride_n = q.stride(0);
+        params.q_stride_h = q.stride(1);
+        params.o_stride_n = o.stride(0);
+        params.o_stride_h = o.stride(1);
+        if (kv_layout == QKVLayout::kNHD) {
+          // (num_pages, page_size, num_heads, head_dim)
+          params.k_stride_n = paged_k_cache.stride(1);
+          params.k_stride_h = paged_k_cache.stride(2);
+          params.v_stride_n = paged_v_cache.stride(1);
+          params.v_stride_h = paged_v_cache.stride(2);
+        } else {
+          // (num_pages, num_heads, page_size, head_dim)
+          params.k_stride_h = paged_k_cache.stride(1);
+          params.k_stride_n = paged_k_cache.stride(2);
+          params.v_stride_h = paged_v_cache.stride(1);
+          params.v_stride_n = paged_v_cache.stride(2);
+        }
+        params.nnz_qo = q.size(0);
+        params.num_qo_heads = q.size(1);
+        params.num_kv_heads = num_kv_heads;
+        params.group_size = params.num_qo_heads / num_kv_heads;
+        params.page_size = page_size;
+        params.window_left = window_left;
+        params.causal = mask_mode_code == 1;
+        params.qo_tile_indices =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.qo_tile_indices_offset);
+        params.qo_indptr = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.qo_indptr_offset);
+        params.kv_indptr = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_indptr_offset);
+        params.qo_lens = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.qo_len_offset);
+        params.kv_lens = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_len_offset);
+        params.head_indices =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.head_indices_offset);
+        params.work_indptr =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.work_indptr_offset);
+        params.kv_indices = static_cast<IdType*>(paged_kv_indices.data_ptr());
+
+        ADDITIONAL_PARAMS_SETTER
+
+        // Not support various head_dim for now
+        static_assert(HEAD_DIM_QK == HEAD_DIM_VO, "head_dim_qk and head_dim_vo should be the same");
+        // Currently only support same quantization precision
+        static_assert(std::is_same_v<DTypeQ, DTypeKV>);
+
+        bool same_schedule_for_all_heads = plan_info.same_schedule_for_all_heads;
+        DISPATCH_BOOL(same_schedule_for_all_heads, SAME_SCHEDULER_FOR_ALL_HEADS, [&] {
+          cudaError_t status =
+              BatchFP8PrefillWithPagedKVCacheDispatched<HEAD_DIM_QK, MASK_MODE, USE_SLIDING_WINDOW,
+                                                        SAME_SCHEDULER_FOR_ALL_HEADS,
+                                                        AttentionVariant>(params, stream);
+          TORCH_CHECK(status == cudaSuccess,
+                      "BatchPrefillWithPagedKVCacheSM90Run failed with error: ",
+                      cudaGetErrorString(status));
+          return true;
+        });
+      });
+}

csrc/single_prefill_fp8_sm90.cu

@@ -0,0 +1,101 @@
+/*
+ * Copyright (c) 2024 by FlashInfer team.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <flashinfer/attention/mask.cuh>
+#include <flashinfer/layout.cuh>
+#include <flashinfer/math.cuh>
+#include <optional>
+
+#include "pytorch_extension_utils.h"
+#include "single_prefill_sm90_config.inc"
+
+namespace flashinfer {
+
+template <uint32_t HEAD_DIM, MaskMode MASK_MODE, bool LEFT_SLINDING_WINDOW,
+          typename AttentionVariant, typename Params>
+cudaError_t SingleFP8PrefillWithKVCacheDispatched(Params& params, cudaStream_t stream);
+
+}  // namespace flashinfer
+
+using namespace flashinfer;
+
+void single_prefill_with_kv_cache_sm90(at::Tensor q, at::Tensor k, at::Tensor v, at::Tensor tmp,
+                                       at::Tensor o, std::optional<at::Tensor> maybe_lse,
+                                       int64_t mask_mode_code, int64_t layout,
+                                       int64_t window_left ADDITIONAL_FUNC_PARAMS) {
+  unsigned int head_dim_qk = q.size(2);
+  unsigned int head_dim_vo = v.size(2);
+  unsigned int num_qo_heads = q.size(1);
+  unsigned int qo_len = q.size(0);
+
+  auto q_scalar_type = q.scalar_type();
+  auto kv_scalar_type = k.scalar_type();
+
+  QKVLayout kv_layout = static_cast<QKVLayout>(layout);
+  const c10::cuda::OptionalCUDAGuard device_guard(q.device());
+  const cudaStream_t stream = c10::cuda::getCurrentCUDAStream();
+  const MaskMode mask_mode = static_cast<MaskMode>(mask_mode_code);
+
+  DISPATCH_context(
+      DTypeQ, DTypeKV, DTypeO, IdType, MASK_MODE, HEAD_DIM_QK, HEAD_DIM_VO, USE_SLIDING_WINDOW,
+      USE_LOGITS_SOFT_CAP, AttentionVariant, Params, [&] {
+        Params params;
+        params.q_ptr = static_cast<DTypeQ*>(q.data_ptr());
+        params.k_ptr = static_cast<DTypeKV*>(k.data_ptr());
+        params.v_ptr = static_cast<DTypeKV*>(v.data_ptr());
+        params.o_ptr = static_cast<DTypeO*>(o.data_ptr());
+        params.lse_ptr = maybe_lse ? (static_cast<float*>(maybe_lse->data_ptr())) : nullptr;
+        params.q_stride_n = q.stride(0);
+        params.q_stride_h = q.stride(1);
+        params.o_stride_n = o.stride(0);
+        params.o_stride_h = o.stride(1);
+        if (kv_layout == QKVLayout::kNHD) {
+          params.k_stride_n = k.stride(0);
+          params.k_stride_h = k.stride(1);
+          params.v_stride_n = v.stride(0);
+          params.v_stride_h = v.stride(1);
+        } else {
+          params.k_stride_h = k.stride(0);
+          params.k_stride_n = k.stride(1);
+          params.v_stride_h = v.stride(0);
+          params.v_stride_n = v.stride(1);
+        }
+        params.qo_len = q.size(0);
+        params.kv_len = k.size(0);
+        params.num_qo_heads = q.size(1);
+        params.num_kv_heads = k.size(1);
+        params.causal = mask_mode == MaskMode::kCausal;
+        params.group_size = params.num_qo_heads / params.num_kv_heads;
+
+        // Note(Yilong): this should be checked on Python Side
+        // Only support window_left == 0 for now
+        params.window_left = window_left;
+
+        // Note(Yilong): all quantization parameters are set in additional_params
+        ADDITIONAL_PARAMS_SETTER
+
+        // Not support various head_dim for now
+        static_assert(HEAD_DIM_QK == HEAD_DIM_VO, "head_dim_qk and head_dim_vo should be the same");
+        // Currently only support same quantization precision
+        static_assert(std::is_same_v<DTypeQ, DTypeKV>);
+
+        cudaError_t status =
+            SingleFP8PrefillWithKVCacheDispatched<HEAD_DIM_QK, MASK_MODE, USE_SLIDING_WINDOW,
+                                                  AttentionVariant>(params, stream);
+        TORCH_CHECK(status == cudaSuccess, "single_prefill_with_kv_cache_sm90 failed with error: " +
+                                               std::string(cudaGetErrorString(status)));
+        return true;
+      });
+}

csrc/single_prefill_fp8_sm90_kernel_inst.jinja

@@ -0,0 +1,11 @@
+#include <flashinfer/attention/hopper/quantization/prefill_sm90.cuh>
+#include "single_prefill_sm90_config.inc"
+
+using namespace flashinfer;
+
+namespace flashinfer {
+
+template cudaError_t SingleFP8PrefillWithKVCacheDispatched
+    <{{ head_dim_qk }}, {{ mask_mode }}, /*USE_SLIDING_WINDOW=*/{{ use_sliding_window }}, {{ variant_name }}, Params>(
+    Params& params, cudaStream_t stream);
+};

flashinfer/decode.py

@@ -530,6 +530,9 @@ def single_decode_with_kv_cache(
             _get_cache_alibi_slopes_buf(num_qo_heads, q.device),
             logits_soft_cap,
             sm_scale,
+            None,  # scale_q, not supported yet
+            None,  # scale_k
+            None,  # scale_v
             rope_scale,
             rope_theta,
         )
@@ -1169,6 +1172,9 @@ def run(
                     _get_cache_alibi_slopes_buf(q.shape[1], q.device),
                     logits_soft_cap,
                     sm_scale,
+                    None,  # scale_q, not supported yet
+                    None,  # scale_k
+                    None,  # scale_v
                     rope_scale,
                     rope_theta,
                 ]