[BugFix] Fix low prediction accuracy of deepseekv3 (#2798)

fastdeploy/model_executor/layers/attention/mla_attention_backend.py

@@ -41,7 +41,8 @@ from fastdeploy.config import FDConfig
 from fastdeploy.model_executor.layers.attention.attention import Attention
 from fastdeploy.model_executor.layers.attention.base_attention_backend import (
     AttentionBackend, AttentionMetadata)
-from fastdeploy.model_executor.layers.attention.utils import init_rank_and_device_id
+from fastdeploy.model_executor.layers.attention.utils import \
+    init_rank_and_device_id
 from fastdeploy.worker.forward_meta import ForwardMeta
 
 
@@ -185,6 +186,8 @@ class MLAAttentionBackend(AttentionBackend):
         # MLA
         metadata.max_enc_len_this_time = metadata.set_max_lengths[1]
         metadata.max_dec_len_this_time = metadata.set_max_lengths[2]
+        forward_meta.max_enc_len_this_time = metadata.set_max_lengths[1]
+        forward_meta.max_dec_len_this_time = metadata.set_max_lengths[2]
 
         # pd_disaggregation
         metadata.kv_signal_data_list = [None] * self.num_layers
@@ -375,9 +378,6 @@ class MLAAttentionBackend(AttentionBackend):
         speculate_decoder = self.speculative_method is not None
         speculate_max_tokens = self.speculate_max_draft_token_num
 
-        decode_stage = forward_meta.is_decode_batch
-        prefill_stage = not (forward_meta.is_decode_batch)
-
         if self.use_pd_disaggregation:
             metadata.kv_signal_data_list[
                 layer.layer_id] = init_signal_layerwise(
@@ -387,8 +387,7 @@ class MLAAttentionBackend(AttentionBackend):
         latent_cache = forward_meta.caches[layer.layer_id] if hasattr(
             forward_meta, 'caches') else None
 
-        if prefill_stage:
+        if k is not None:
-            # 写入缓存
             prefill_mla_write_cache(
                 compressed_kv,
                 k_pe,
@@ -419,8 +418,7 @@ class MLAAttentionBackend(AttentionBackend):
             return fmha_out
 
         # Decode
-        if decode_stage:
+        if k is None:
-            # mla写入缓存
             decode_mla_write_cache(
                 compressed_kv,
                 k_pe,

fastdeploy/model_executor/models/deepseek_v3.py

@@ -317,10 +317,7 @@ class DeepseekV3MLAAttention(nn.Layer):
         ],
                                 dtype=layernorm_out.dtype)
 
-        decode_stage = forward_meta.is_decode_batch
+        if forward_meta.max_enc_len_this_time:
-        prefill_stage = not (forward_meta.is_decode_batch)
-
-        if prefill_stage:
             query = self.q_a_proj(layernorm_out)
             query = self.q_a_layernorm(query)
             query = self.q_b_proj(query)
@@ -370,8 +367,7 @@ class DeepseekV3MLAAttention(nn.Layer):
                 fmha_out_prefill.dtype)
 
             fmha_out = fmha_out + fmha_out_prefill
-
+        if forward_meta.max_dec_len_this_time:
-        if decode_stage:
             query = self.q_a_proj(layernorm_out)
             query = self.q_a_layernorm(query)
             ln_out_or_q_c = query
@@ -554,28 +550,6 @@ class DeepSeekV3Model(nn.Layer):
             prefix="deepseek_v3.norm",
         )
 
-    def pre_process(self, forward_meta):
-        """
-        """
-        seq_lens_encoder = forward_meta.seq_lens_encoder
-        seq_lens_decoder = forward_meta.seq_lens_decoder
-        seq_lens_this_time = forward_meta.seq_lens_this_time
-        position_ids_shape = paddle.sum(seq_lens_this_time)
-
-        position_ids = paddle.empty(shape=position_ids_shape,
-                                    dtype=seq_lens_encoder.dtype)
-        mask_encoder_batch = paddle.empty(
-            shape=position_ids_shape,
-            dtype=seq_lens_encoder.dtype).unsqueeze(1)
-
-        get_position_ids_and_mask_encoder_batch(seq_lens_encoder,
-                                                seq_lens_decoder,
-                                                seq_lens_this_time,
-                                                position_ids,
-                                                mask_encoder_batch)
-
-        return position_ids, mask_encoder_batch
-
     def load_state_dict(self, state_dict):
         """
         Load model parameters from a given state dictionary.
@@ -590,13 +564,13 @@ class DeepSeekV3Model(nn.Layer):
         self,
         ids_remove_padding: paddle.Tensor,
         forward_meta: ForwardMeta,
+        position_ids: paddle.Tensor,
+        mask_encoder_batch: paddle.Tensor,
     ):
         """
         """
         hidden_states = self.embeddings(ids_remove_padding=ids_remove_padding)
 
-        position_ids, mask_encoder_batch = self.pre_process(forward_meta)
-
         residual = None
         for i in range(self.num_layers):
             hidden_states, residual = self.decoder_layers[i](
@@ -650,6 +624,27 @@ class DeepseekV3ForCausalLM(ModelForCasualLM):
         logits[:, self.ori_vocab_size:] = -float("inf")
         return logits
 
+    def pre_process(self, forward_meta):
+        """
+        """
+        seq_lens_encoder = forward_meta.seq_lens_encoder
+        seq_lens_decoder = forward_meta.seq_lens_decoder
+        seq_lens_this_time = forward_meta.seq_lens_this_time
+        position_ids_shape = paddle.sum(seq_lens_this_time)
+        position_ids = paddle.empty(shape=position_ids_shape,
+                                    dtype=seq_lens_encoder.dtype)
+        mask_encoder_batch = paddle.empty(
+            shape=position_ids_shape,
+            dtype=seq_lens_encoder.dtype).unsqueeze(1)
+
+        get_position_ids_and_mask_encoder_batch(seq_lens_encoder,
+                                                seq_lens_decoder,
+                                                seq_lens_this_time,
+                                                position_ids,
+                                                mask_encoder_batch)
+
+        return position_ids, mask_encoder_batch
+
     def forward(
         self,
         ids_remove_padding: paddle.Tensor,
@@ -657,7 +652,9 @@ class DeepseekV3ForCausalLM(ModelForCasualLM):
     ):
         """
         """
-        hidden_states = self.model(ids_remove_padding, forward_meta)
+        position_ids, mask_encoder_batch = self.pre_process(forward_meta)
+        hidden_states = self.model(ids_remove_padding=ids_remove_padding, forward_meta=forward_meta,
+            position_ids=position_ids, mask_encoder_batch=mask_encoder_batch)
         return hidden_states