[FP8][Kernel] Dynamic kv cache scaling factors computation (#11906)

Signed-off-by: Gregory Shtrasberg <[email protected]>
Co-authored-by: Micah Williamson <[email protected]>

benchmarks/kernels/benchmark_paged_attention.py

@@ -98,7 +98,9 @@ def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
         start_time = time.perf_counter()
 
         # Using default kv_scale
-        k_scale = v_scale = 1.0
+        k_scale = v_scale = torch.tensor(1.0,
+                                         dtype=torch.float32,
+                                         device=device)
 
         for _ in range(num_iters):
             if version == "v1":

csrc/attention/attention_kernels.cuh

@@ -105,7 +105,7 @@ __device__ void paged_attention_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float k_scale, const float v_scale, const int tp_rank,
+    const float* k_scale, const float* v_scale, const int tp_rank,
     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   const int seq_idx = blockIdx.y;
@@ -285,7 +285,7 @@ __device__ void paged_attention_kernel(
           Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(
               k_ptr + offset1 * BLOCK_SIZE * x + offset2);
           k_vecs[j] = fp8::scaled_convert<K_vec, Quant_vec, KV_DTYPE>(
-              k_vec_quant, k_scale);
+              k_vec_quant, *k_scale);
         }
       }
 
@@ -415,7 +415,7 @@ __device__ void paged_attention_kernel(
               *reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
           // Vector conversion from V_quant_vec to V_vec.
           v_vec = fp8::scaled_convert<V_vec, V_quant_vec, KV_DTYPE>(v_quant_vec,
-                                                                    v_scale);
+                                                                    *v_scale);
         }
         if (block_idx == num_seq_blocks - 1) {
           // NOTE(woosuk): When v_vec contains the tokens that are out of the
@@ -513,7 +513,7 @@ __global__ void paged_attention_v1_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float k_scale, const float v_scale, const int tp_rank,
+    const float* k_scale, const float* v_scale, const int tp_rank,
     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
@@ -549,7 +549,7 @@ __global__ void paged_attention_v2_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float k_scale, const float v_scale, const int tp_rank,
+    const float* k_scale, const float* v_scale, const int tp_rank,
     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,

csrc/attention/paged_attention_v1.cu

@@ -41,7 +41,7 @@
           out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, \
           scale, block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq,    \
           alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride,      \
-          k_scale, v_scale, tp_rank, blocksparse_local_blocks,              \
+          k_scale_ptr, v_scale_ptr, tp_rank, blocksparse_local_blocks,      \
           blocksparse_vert_stride, blocksparse_block_size,                  \
           blocksparse_head_sliding_step);
 
@@ -53,10 +53,10 @@ void paged_attention_v1_launcher(
     torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, int num_kv_heads, float scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
-    const std::optional<torch::Tensor>& alibi_slopes, float k_scale,
-    float v_scale, const int tp_rank, const int blocksparse_local_blocks,
-    const int blocksparse_vert_stride, const int blocksparse_block_size,
-    const int blocksparse_head_sliding_step) {
+    const std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   int num_seqs = query.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
@@ -80,6 +80,8 @@ void paged_attention_v1_launcher(
   CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
   int* block_tables_ptr = block_tables.data_ptr<int>();
   int* seq_lens_ptr = seq_lens.data_ptr<int>();
+  const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
+  const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
 
   constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
   int padded_max_seq_len =
@@ -177,8 +179,9 @@ void paged_attention_v1(
     torch::Tensor& seq_lens,      // [num_seqs]
     int64_t block_size, int64_t max_seq_len,
     const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);

csrc/attention/paged_attention_v2.cu

@@ -37,7 +37,7 @@
           exp_sums_ptr, max_logits_ptr, tmp_out_ptr, query_ptr, key_cache_ptr, \
           value_cache_ptr, num_kv_heads, scale, block_tables_ptr,              \
           seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride,    \
-          kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,          \
+          kv_block_stride, kv_head_stride, k_scale_ptr, v_scale_ptr, tp_rank,  \
           blocksparse_local_blocks, blocksparse_vert_stride,                   \
           blocksparse_block_size, blocksparse_head_sliding_step);              \
   vllm::paged_attention_v2_reduce_kernel<T, HEAD_SIZE, NUM_THREADS,            \
@@ -54,10 +54,10 @@ void paged_attention_v2_launcher(
     torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, int num_kv_heads, float scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
-    const std::optional<torch::Tensor>& alibi_slopes, float k_scale,
-    float v_scale, const int tp_rank, const int blocksparse_local_blocks,
-    const int blocksparse_vert_stride, const int blocksparse_block_size,
-    const int blocksparse_head_sliding_step) {
+    const std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   int num_seqs = query.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
@@ -84,6 +84,8 @@ void paged_attention_v2_launcher(
   CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
   int* block_tables_ptr = block_tables.data_ptr<int>();
   int* seq_lens_ptr = seq_lens.data_ptr<int>();
+  const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
+  const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
 
   constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
   int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
@@ -188,8 +190,9 @@ void paged_attention_v2(
     torch::Tensor& seq_lens,      // [num_seqs]
     int64_t block_size, int64_t max_seq_len,
     const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);

csrc/cache.h

@@ -18,15 +18,15 @@ void copy_blocks(std::vector<torch::Tensor> const& key_caches,
 void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
                        torch::Tensor& key_cache, torch::Tensor& value_cache,
                        torch::Tensor& slot_mapping,
-                       const std::string& kv_cache_dtype, const double k_scale,
-                       const double v_scale);
+                       const std::string& kv_cache_dtype,
+                       torch::Tensor& k_scale, torch::Tensor& v_scale);
 
 void reshape_and_cache_flash(torch::Tensor& key, torch::Tensor& value,
                              torch::Tensor& key_cache,
                              torch::Tensor& value_cache,
                              torch::Tensor& slot_mapping,
                              const std::string& kv_cache_dtype,
-                             const double k_scale, const double v_scale);
+                             torch::Tensor& k_scale, torch::Tensor& v_scale);
 
 // Just for unittest
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,

csrc/cache_kernels.cu

@@ -159,8 +159,8 @@ __global__ void reshape_and_cache_kernel(
                                          // block_size]
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int key_stride, const int value_stride, const int num_heads,
-    const int head_size, const int block_size, const int x, const float k_scale,
-    const float v_scale) {
+    const int head_size, const int block_size, const int x,
+    const float* k_scale, const float* v_scale) {
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   if (slot_idx < 0) {
@@ -196,9 +196,9 @@ __global__ void reshape_and_cache_kernel(
       value_cache[tgt_value_idx] = tgt_value;
     } else {
       key_cache[tgt_key_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, k_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
       value_cache[tgt_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, v_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
     }
   }
 }
@@ -214,7 +214,7 @@ __global__ void reshape_and_cache_flash_kernel(
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int block_stride, const int key_stride, const int value_stride,
     const int num_heads, const int head_size, const int block_size,
-    const float k_scale, const float v_scale) {
+    const float* k_scale, const float* v_scale) {
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   // NOTE: slot_idx can be -1 if the token is padded
@@ -239,9 +239,9 @@ __global__ void reshape_and_cache_flash_kernel(
       value_cache[tgt_key_value_idx] = tgt_value;
     } else {
       key_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, k_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
       value_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, v_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
     }
   }
 }
@@ -258,7 +258,9 @@ __global__ void reshape_and_cache_flash_kernel(
           reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),           \
           reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),         \
           slot_mapping.data_ptr<int64_t>(), key_stride, value_stride, \
-          num_heads, head_size, block_size, x, k_scale, v_scale);
+          num_heads, head_size, block_size, x,                        \
+          reinterpret_cast<const float*>(k_scale.data_ptr()),         \
+          reinterpret_cast<const float*>(v_scale.data_ptr()));
 
 void reshape_and_cache(
     torch::Tensor& key,    // [num_tokens, num_heads, head_size]
@@ -268,8 +270,8 @@ void reshape_and_cache(
     torch::Tensor&
         value_cache,  // [num_blocks, num_heads, head_size, block_size]
     torch::Tensor& slot_mapping,  // [num_tokens]
-    const std::string& kv_cache_dtype, const double k_scale,
-    const double v_scale) {
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale) {
   int num_tokens = key.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);
@@ -299,7 +301,9 @@ void reshape_and_cache(
           reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),           \
           reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),         \
           slot_mapping.data_ptr<int64_t>(), block_stride, key_stride, \
-          value_stride, num_heads, head_size, block_size, k_scale, v_scale);
+          value_stride, num_heads, head_size, block_size,             \
+          reinterpret_cast<const float*>(k_scale.data_ptr()),         \
+          reinterpret_cast<const float*>(v_scale.data_ptr()));
 
 void reshape_and_cache_flash(
     torch::Tensor& key,        // [num_tokens, num_heads, head_size]
@@ -308,8 +312,8 @@ void reshape_and_cache_flash(
     torch::Tensor&
         value_cache,  // [num_blocks, block_size, num_heads, head_size]
     torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
-    const std::string& kv_cache_dtype, const double k_scale,
-    const double v_scale) {
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale) {
   // NOTE(woosuk): In vLLM V1, key.size(0) can be different from
   // slot_mapping.size(0) because of padding for CUDA graphs.
   // In vLLM V0, key.size(0) is always equal to slot_mapping.size(0) because

csrc/cpu/attention.cpp

@@ -460,11 +460,11 @@ void paged_attention_v1(
     torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
     int64_t max_seq_len, const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step) {
-  TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f);
   TORCH_CHECK(blocksparse_vert_stride <= 1,
               "CPU backend does not support blocksparse attention yet.");
   VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v1_impl",
@@ -782,11 +782,11 @@ void paged_attention_v2(
     torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
     int64_t max_seq_len, const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step) {
-  TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f);
   TORCH_CHECK(blocksparse_vert_stride <= 1,
               "CPU backend does not support blocksparse attention yet.");
   VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v2_impl",

csrc/cpu/cache.cpp

@@ -107,10 +107,8 @@ void copy_blocks(std::vector<torch::Tensor> const& key_caches,
 void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
                        torch::Tensor& key_cache, torch::Tensor& value_cache,
                        torch::Tensor& slot_mapping,
-                       const std::string& kv_cache_dtype, double k_scale,
-                       double v_scale) {
-  TORCH_CHECK(k_scale == 1.0f && v_scale == 1.0f);
-
+                       const std::string& kv_cache_dtype,
+                       torch::Tensor& k_scale, torch::Tensor& v_scale) {
   int num_tokens = key.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);

csrc/cpu/torch_bindings.cpp

@@ -30,7 +30,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "    Tensor value_cache, int num_kv_heads, float scale,"
       "    Tensor block_tables, Tensor seq_lens, int block_size,"
       "    int max_seq_len, Tensor? alibi_slopes,"
-      "    str kv_cache_dtype, float k_scale, float v_scale,"
+      "    str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
       "    int tp_rank, int blocksparse_local_blocks,"
       "    int blocksparse_vert_stride, int blocksparse_block_size,"
       "    int blocksparse_head_sliding_step) -> ()");
@@ -44,7 +44,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "    Tensor value_cache, int num_kv_heads, float scale,"
       "    Tensor block_tables, Tensor seq_lens, int block_size,"
       "    int max_seq_len, Tensor? alibi_slopes,"
-      "    str kv_cache_dtype, float k_scale, float v_scale,"
+      "    str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
       "    int tp_rank, int blocksparse_local_blocks,"
       "    int blocksparse_vert_stride, int blocksparse_block_size,"
       "    int blocksparse_head_sliding_step) -> ()");
@@ -148,7 +148,7 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
       "                  Tensor! key_cache, Tensor! value_cache,"
       "                  Tensor slot_mapping,"
       "                  str kv_cache_dtype,"
-      "                  float k_scale, float v_scale) -> ()");
+      "                  Tensor k_scale, Tensor v_scale) -> ()");
   cache_ops.impl("reshape_and_cache", torch::kCPU, &reshape_and_cache);
 }
 

csrc/ops.h

@@ -34,8 +34,9 @@ void paged_attention_v1(
     torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
     int64_t max_seq_len, const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step);
 
@@ -45,8 +46,9 @@ void paged_attention_v2(
     torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
     int64_t max_seq_len, const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, double k_scale, double v_scale,
-    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step);