fix dsk diff

Author: yuanlehome

csrc/gpu/step.cu

@@ -31,6 +31,7 @@ __global__ void free_and_dispatch_block(bool *stop_flags,
                                         int *used_list_len,
                                         int *free_list,
                                         int *free_list_len,
+                                        int64_t *first_token_ids,
                                         const int bsz,
                                         const int block_size,
                                         const int block_num_per_seq,
@@ -43,6 +44,7 @@ __global__ void free_and_dispatch_block(bool *stop_flags,
         int *block_table_now = block_tables + tid * block_num_per_seq;
         if (stop_flags[tid] && !is_block_step[tid]) {
             // 回收block块
+            first_token_ids[tid] = -1;
             const int encoder_block_len = encoder_block_lens[tid];
             const int decoder_used_len = used_list_len[tid];
             if (decoder_used_len > 0) {
@@ -166,11 +168,11 @@ __global__ void recover_block(int *recover_block_list, // [bsz]
                               int *encoder_block_lens,
                               int *used_list_len,
                               const int64_t *next_tokens,
+                              const int64_t *first_token_ids,
                               const int bsz,
                               const int block_num_per_seq,
                               const int length,
-                              const int pre_id_length,
-                              const int first_token_id) {
+                              const int pre_id_length) {
     const int bid = blockIdx.x;
     const int tid = threadIdx.x;
     __shared__ int ori_free_list_len;
@@ -189,7 +191,8 @@ __global__ void recover_block(int *recover_block_list, // [bsz]
             seq_lens_encoder[recover_id] = seq_len;
             stop_flags[recover_id] = false;
             input_ids_now[ori_seq_len_encoder + step_idx_now - 1] = next_tokens[recover_id]; // next tokens
-            input_ids_now[0] = first_token_id; // set first prompt token
+            input_ids_now[0] =
+                first_token_ids[recover_id];  // set first prompt token
             const int ori_free_list_len_tid0 = atomicSub(free_list_len, decoder_used_len);
             ori_free_list_len = ori_free_list_len_tid0;
 #ifdef DEBUG_STEP
@@ -234,9 +237,9 @@ void StepPaddle(const paddle::Tensor& stop_flags,
                 const paddle::Tensor& pre_ids,
                 const paddle::Tensor& step_idx,
                 const paddle::Tensor& next_tokens,
+                const paddle::Tensor &first_token_ids,
                 const int block_size,
                 const int encoder_decoder_block_num,
-                const int64_t first_token_id,
                 const int speculate_step_token_num) {
     auto cu_stream = seq_lens_this_time.stream();
     const int bsz = seq_lens_this_time.shape()[0];
@@ -264,6 +267,7 @@ void StepPaddle(const paddle::Tensor& stop_flags,
         const_cast<int*>(used_list_len.data<int>()),
         const_cast<int*>(free_list.data<int>()),
         const_cast<int*>(free_list_len.data<int>()),
+        const_cast<int64_t *>(first_token_ids.data<int64_t>()),
         bsz,
         block_size,
         block_num_per_seq,
@@ -300,11 +304,11 @@ void StepPaddle(const paddle::Tensor& stop_flags,
             const_cast<int*>(encoder_block_lens.data<int>()),
             const_cast<int*>(used_list_len.data<int>()),
             next_tokens.data<int64_t>(),
+            first_token_ids.data<int64_t>(),
             bsz,
             block_num_per_seq,
             length,
-            pre_id_length,
-            first_token_id
+            pre_id_length
         );
 #ifdef DEBUG_STEP
 #ifdef PADDLE_WITH_HIP
@@ -337,10 +341,10 @@ PD_BUILD_OP(step_paddle)
              "input_ids",
              "pre_ids",
              "step_idx",
-             "next_tokens"})
+             "next_tokens",
+             "first_token_ids",})
     .Attrs({"block_size: int",
             "encoder_decoder_block_num: int",
-            "first_token_id: int64_t",
             "speculate_step_token_num: int"})
     .Outputs({"stop_flags_out",
               "seq_lens_this_time_out",
@@ -358,7 +362,8 @@ PD_BUILD_OP(step_paddle)
               "used_list_len_out",
               "free_list_out",
               "free_list_len_out",
-              "input_ids_out"})
+              "input_ids_out",
+              "first_token_ids_out",})
     .SetInplaceMap({{"stop_flags", "stop_flags_out"},
                     {"seq_lens_this_time", "seq_lens_this_time_out"},
                     {"seq_lens_encoder", "seq_lens_encoder_out"},
@@ -375,5 +380,6 @@ PD_BUILD_OP(step_paddle)
                     {"used_list_len", "used_list_len_out"},
                     {"free_list", "free_list_out"},
                     {"free_list_len", "free_list_len_out"},
-                    {"input_ids", "input_ids_out"}})
+                    {"input_ids", "input_ids_out"},
+                    {"first_token_ids", "first_token_ids_out"}})
     .SetKernelFn(PD_KERNEL(StepPaddle));

paddlenlp/experimental/transformers/deepseek_v2/modeling.py

@@ -92,10 +92,12 @@ def __init__(
             * attn_factor
         )
 
-        cache = self._compute_cos_sin_cache()
+        cos_cache, sin_cache = self._compute_cos_sin_cache()
 
-        self.cos_sin_cache: paddle.Tensor
-        self.register_buffer("cos_sin_cache", cache, persistable=True)
+        self.cos_cache: paddle.Tensor
+        self.register_buffer("cos_cache", cos_cache, persistable=True)
+        self.sin_cache: paddle.Tensor
+        self.register_buffer("sin_cache", sin_cache, persistable=True)
 
     def _compute_inv_freq(self, scaling_factor: float) -> paddle.Tensor:
         pos_freqs = self.base ** (paddle.arange(0, self.rotary_dim, 2, dtype=paddle.float32) / self.rotary_dim)
@@ -114,23 +116,37 @@ def _compute_inv_freq(self, scaling_factor: float) -> paddle.Tensor:
     def _compute_cos_sin_cache(self) -> paddle.Tensor:
         inv_freq = self._compute_inv_freq(self.scaling_factor)
         t = paddle.arange(self.max_position_embeddings * self.scaling_factor, dtype=paddle.float32)
-        freqs = paddle.einsum("i,j->ij", t, inv_freq)
-        cos = freqs.cos() * self.mscale
-        sin = freqs.sin() * self.mscale
-        cache = paddle.concat((cos, sin), axis=-1)
-        return cache.cast(self._dtype)
+
+        freqs = paddle.outer(t, inv_freq)
+        emb = paddle.concat((freqs, freqs), axis=-1)
+        cos = emb.cos() * self.mscale
+        sin = emb.sin() * self.mscale
+
+        return cos.cast(self._dtype) , sin.cast(self._dtype)
 
     def forward(
         self,
         position_ids: paddle.Tensor,
         query: paddle.Tensor,
         key: paddle.Tensor,
     ) -> Tuple[paddle.Tensor, paddle.Tensor]:
-        from paddlenlp_ops import fused_rotary_position_encoding
+        cos = self.cos_cache[position_ids].unsqueeze(1)
+        sin = self.sin_cache[position_ids].unsqueeze(1)
+
+        def rotate_half(x):
+            """Rotates half the hidden axiss of the input."""
+            x1 = x[..., : x.shape[-1] // 2]
+            x2 = x[..., x.shape[-1] // 2 :]
+            return paddle.concat([-x2, x1], axis=-1)  # shape is the same as x
+
+        s, h, d = query.shape
+        query = query.reshape([s, h, d // 2, 2]).transpose([0, 1, 3, 2]).reshape([s, h, d])
+
+        s, h, d = key.shape
+        key = key.reshape([s, h, d // 2, 2]).transpose([0, 1, 3, 2]).reshape([s, h, d])
 
-        # In-place operations that update the query and key tensors.
-        os.environ["stride_in_no_check_dy2st_diff"] = "1"
-        fused_rotary_position_encoding(query, key, position_ids, self.cos_sin_cache, self.rotary_dim, False)
+        query = (query * cos) + (rotate_half(query) * sin)
+        key = (key * cos) + (rotate_half(key) * sin)
 
         return query, key
 

paddlenlp/experimental/transformers/fused_transformer_layers.py

@@ -1160,7 +1160,7 @@ def compute_fused_moe(self, tmp_out, i):
         def get_moe_scores(
             gating_output: paddle.Tensor,
             config: MoeConfig,
-        ) -> (paddle.Tensor, paddle.Tensor):
+        ) -> tuple[paddle.Tensor, paddle.Tensor]:
 
             num_token = gating_output.shape[0]
             num_expert_group = config.num_expert_group