fix

flexflow · Oct 20, 2024 · 9093024 · 9093024
1 parent 222ce02
commit 9093024
Show file tree

Hide file tree

Showing 3 changed files with 28 additions and 15 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -240,7 +240,7 @@ if(NOT BUILD_LEGION_ONLY)
     list(APPEND FF_CC_FLAGS
       -DFF_DEBUG)
     list(APPEND FF_NVCC_FLAGS
-      -DFF_DEBUG)
+      -DFF_DEBUG -lineinfo)
   endif()
 
   message(STATUS "FlexFlow MAX_DIM: ${FF_MAX_DIM}")

diff --git a/src/c/flexflow_c.cc b/src/c/flexflow_c.cc
@@ -1730,7 +1730,7 @@ void flexflow_model_generate(flexflow_model_t handle_,
                   handle,
                   dataset_fp.c_str(),
                   max_lengths[i],
-                  max_new_tokens[i],
+                  max_new_tokens_[i],
                   training_steps[i]);
     } else {
       assert(false && "Unknown request type");

diff --git a/src/ops/inc_multihead_self_attention.cu b/src/ops/inc_multihead_self_attention.cu
@@ -384,7 +384,8 @@ __global__ void compute_attention_kernel_generation_kernel(
     int max_seq_length,
     int per_head_size,
     int hidden_size,
-    BatchConfig::PerRequestInfo *request_infos) {
+    BatchConfig::PerRequestInfo *request_infos,
+    BatchConfig::PerTokenInfo *tokens_infos) {
 
   // q, k
   using Q_vec = typename VEC_K<DT, THREADS_PER_KEY>::Type;
@@ -409,11 +410,14 @@ __global__ void compute_attention_kernel_generation_kernel(
   int const tidx = threadIdx.x;
   // head id
   int const head_idx = blockIdx.x;
-  // request idx
-  int const request_idx = blockIdx.y;
-
+  // generation token idx.
+  int const batch_config_token_idx = blockIdx.y;
   int const batch_config_request_id =
-      request_infos[request_idx].batch_config_request_id;
+      tokens_infos[batch_config_token_idx].request_index;
+  if (request_infos[batch_config_request_id].prompt_phase) {
+    return;
+  }
+  assert(request_infos[batch_config_request_id].num_tokens_in_batch == 1);
 
   int const first_step = 0;
 
@@ -432,11 +436,13 @@ __global__ void compute_attention_kernel_generation_kernel(
   // first WARPS_PER_BLOCK for store qk_max, second WARPS_PER_BLOCK for sum
   __shared__ float red_smem[WARPS_PER_BLOCK * 2];
 
-  const DT *q_ptr = query + request_idx * hidden_size * QKV_WEIGHT_NUM +
+  const DT *q_ptr = query +
+                    batch_config_token_idx * hidden_size * QKV_WEIGHT_NUM +
                     head_idx * per_head_size;
   __shared__ Q_vec q_vecs[THREADS_PER_KEY][K_VECS_PER_THREAD];
   // DT const *q_ptr =
-  //     query + request_idx * Dh * QKV_WEIGHT_NUM + head_idx * per_head_size;
+  //     query + batch_config_token_idx * Dh * QKV_WEIGHT_NUM + head_idx *
+  //     per_head_size;
 
   // q tensor in this thread
   // if THREADS_PER_KEY is 4, first thread load 0, 4, 8, 12..., total
@@ -615,10 +621,10 @@ __global__ void compute_attention_kernel_generation_kernel(
 
   // Output the final values.
   if (vo == 0 && (Dh == Dh_MAX || vi < Dh)) {
-    convert_from_float(
-        *reinterpret_cast<V_vec *>(output_ptr + request_idx * hidden_size +
-                                   head_idx * per_head_size + vi),
-        out);
+    convert_from_float(*reinterpret_cast<V_vec *>(
+                           output_ptr + batch_config_token_idx * hidden_size +
+                           head_idx * per_head_size + vi),
+                       out);
   }
 }
 
@@ -897,14 +903,15 @@ void update_kv_cache_kernel(IncMultiHeadSelfAttentionMeta const *m,
           BatchConfig::max_sequence_length(),                                  \
           m->qProjSize,                                                        \
           m->hidden_size,                                                      \
-          m->request_infos)
+          m->request_infos,                                                    \
+          m->token_infos)
 
 template <typename DT>
 void compute_attention_kernel_generation(IncMultiHeadSelfAttentionMeta const *m,
                                          BatchConfig const *bc,
                                          DT *output_ptr,
                                          cudaStream_t stream) {
-  dim3 grid(m->num_q_heads, bc->num_generation_tokens);
+  dim3 grid(m->num_q_heads, bc->num_active_tokens());
   int const per_head_size = m->qProjSize;
   float scale = (*m->qk_prod_scaling) ? 1.0f / sqrt(m->kProjSize) : 1.0f;
   size_t smem_sz;
@@ -958,8 +965,14 @@ void inference_kernel(IncMultiHeadSelfAttentionMeta *m,
   // phase 1: Implement kernel to apply rotary embedding and scaling
   compute_qkv_kernel(
       m, bc, shard_id, static_cast<DT *>(m->devQKVProjArray), stream);
+  // key/value cache shape (keyed by request idx in batch config):
+  // [proj_size * num_heads, max_seq_length, max_requests]
   update_kv_cache_kernel<DT>(m, bc, stream);
 
+  // QK prods shape (single request only): [num_new_tokens, total_tokens,
+  // num_heads] Attn heads shape (single request only): [vProjSize, num_heads,
+  // num_new_tokens]
+
   if (bc->num_generation_tokens > 0) {
     // phase 3: Compute attention score for generation tokens
     compute_attention_kernel_generation<DT>(