Upload modeling_hunyuan.py

modeling_hunyuan.py

@@ -1,16 +1,5 @@
-# Copyright (C) 2024 THL A29 Limited, a Tencent company.  All rights reserved.
-#
-# Licensed under the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     https://github.com/Tencent/Tencent-Hunyuan-Large/blob/main/License.docx
-#
-# 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.
+# coding=utf-8
+# Copyright 2024 Tencent Inc. All Rights Reserved.
 #
 """ PyTorch HunYuan model."""
 
@@ -74,8 +63,7 @@ _CONFIG_FOR_DOC = "HunYuanConfig"
 def topkgating(logits: Tensor, topk: int):
     logits = logits.float()
     gates = F.softmax(logits, dim=1)
-    # expert_capacity = topk * gates.shape[0]
-    expert_capacity = max(topk, topk * gates.shape[0] // gates.shape[1])
+    expert_capacity = topk * gates.shape[0]
     num_experts = int(gates.shape[1])
     # Top-k router probability and corresponding expert indices for each token.
     # Shape: [tokens_per_group, num_selected_experts].
@@ -265,7 +253,7 @@ class HunYuanRotaryEmbedding(nn.Module):
         self.max_position_embeddings = max_position_embeddings
         self.base = base
         inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
-        # inv_freq = inv_freq.bfloat16()
+        inv_freq = inv_freq.bfloat16()
         self.register_buffer("inv_freq", inv_freq, persistent=False)
 
         # Build here to make `torch.jit.trace` work.
@@ -277,7 +265,6 @@ class HunYuanRotaryEmbedding(nn.Module):
         self.max_seq_len_cached = seq_len
         t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.float32)
 
-        self.inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
         freqs = torch.outer(t, self.inv_freq)
         # Different from paper, but it uses a different permutation in order to obtain the same calculation
         emb = torch.cat((freqs, freqs), dim=-1).float()
@@ -286,7 +273,7 @@ class HunYuanRotaryEmbedding(nn.Module):
 
     def forward(self, x, seq_len=None):
         # x: [bs, num_attention_heads, seq_len, head_size]
-        if seq_len > self.max_seq_len_cached or self.inv_freq.dtype != torch.float32:
+        if seq_len > self.max_seq_len_cached:
             self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
 
         return (
@@ -411,7 +398,7 @@ class HunYuanMLP(nn.Module):
         self.layer_idx = layer_idx
         self.hidden_size = config.hidden_size
         if is_shared_mlp:
-            self.intermediate_size = config.intermediate_size * config.num_shared_expert[0]
+            self.intermediate_size = config.intermediate_size * config.num_shared_expert
         else:
             self.intermediate_size = config.intermediate_size
         self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
@@ -462,66 +449,142 @@ class HunYuanTopKGate(nn.Module):
         if self.moe_topk == 1:
             gate_output = top1gating(logits, random_routing_dropped_token=self.random_routing_dropped_token)
         else:
-            gate_output = topkgating(logits, self.moe_topk[0])
+            gate_output = topkgating(logits, self.moe_topk)
 
         return gate_output
 
 
 class HunYuanMoE(nn.Module):
+    """Mixture-of-Experts block with vectorized expert execution and Straight‑Through Estimator (STE) utilities.
+
+    This implementation removes all Python‑side loops over experts. Expert parameters are **stacked** on‑the‑fly and
+    all experts are executed in a single batched matmul sequence, which allows efficient tensor‑parallel execution
+    (e.g. with DeepSpeed ZeRO‑3) while keeping the state‑dict format unchanged (each expert remains an individual
+    sub‑module for full compatibility with existing checkpoints)."""
+
     def __init__(self, config: HunYuanConfig, layer_idx: Optional[int] = None):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
         self.moe_topk = config.moe_topk
         self.num_experts = config.num_experts
+
+        # Optional shared MLP branch (mixed MoE + dense)
         if config.use_mixed_mlp_moe:
             self.shared_mlp = HunYuanMLP(config, layer_idx=layer_idx, is_shared_mlp=True)
+
+        # Router
         self.gate = HunYuanTopKGate(config, layer_idx=layer_idx)
+
+        # Experts kept as individual sub‑modules so that load_state_dict / save_pretrained stay identical.
         self.experts = nn.ModuleList(
             [HunYuanMLP(config, layer_idx=layer_idx, is_shared_mlp=False) for _ in range(config.num_experts)]
         )
 
-    def forward(self, hidden_states):
+    # ---------------------------------------------------------------------
+    # Internal helpers
+    # ---------------------------------------------------------------------
+    def _stack_weights(self):
+        """Return stacked (batched) expert weights for gate/up/down projections.
+
+        Shapes:
+            Wg : (E, I, H)
+            Wu : (E, I, H)
+            Wd : (E, H, I)  – note transposed compared to nn.Linear's weight so that
+                              torch.matmul(act, Wd.transpose(-2,‑1)) produces (E, C, H)
+        """
+        Wg = torch.stack([exp.gate_proj.weight for exp in self.experts], dim=0)
+        Wu = torch.stack([exp.up_proj.weight for exp in self.experts], dim=0)
+        Wd = torch.stack([exp.down_proj.weight for exp in self.experts], dim=0)
+        return Wg, Wu, Wd
+
+    # ---------------------------------------------------------------------
+    # Public API
+    # ---------------------------------------------------------------------
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        *,
+        return_router_logits: bool = False,
+    ):
+        """Sparse Top‑k MoE forward pass ("y_sparse") with optional router logits.
+
+        This is the route used during both training and inference for the *sparse*
+        branch. No Python loops over experts are used."
+        """
         bsz, seq_len, hidden_size = hidden_states.shape
 
+        # Optional dense branch when using mixed MoE
         if self.config.use_mixed_mlp_moe:
             hidden_states_mlp = self.shared_mlp(hidden_states)
 
-        l_moe, combine_weights, dispatch_mask, exp_counts = self.gate(hidden_states)
-
-        reshaped_input = hidden_states.reshape(-1, hidden_size)
+        # ---------------- Routing ----------------
+        l_moe, combine_weights, dispatch_mask, _ = self.gate(hidden_states)
+
+        flat_input = hidden_states.reshape(-1, hidden_size)                        # (S, H) where S = B*T
+        # dispatch tokens → experts
+        dispatched_input = torch.einsum("sec,sm->ecm",                            # (E, C, H)
+                                        dispatch_mask.to(hidden_states.dtype),
+                                        flat_input)
+
+        # ---------------- Expert computation ----------------
+        Wg, Wu, Wd = self._stack_weights()
+        Wg = Wg.to(hidden_states.dtype)
+        Wu = Wu.to(hidden_states.dtype)
+        Wd = Wd.to(hidden_states.dtype)
+
+        gate_out = torch.einsum("ech,eih->eci", dispatched_input, Wg)             # (E, C, I)
+        up_out   = torch.einsum("ech,eih->eci", dispatched_input, Wu)             # (E, C, I)
+        act_fn   = self.experts[0].act_fn
+        interm   = act_fn(gate_out) * up_out                                      # (E, C, I)
+        expert_output = torch.matmul(interm, Wd.transpose(-2, -1))                # (E, C, H)
+
+        # ---------------- Combine ----------------
+        combined_output = torch.einsum("sec,ecm->sm",                             # (S, H)
+                                       combine_weights.to(hidden_states.dtype),
+                                       expert_output)
+        combined_output = combined_output.reshape(bsz, seq_len, hidden_size)
 
-        dispatched_input = torch.einsum("sec,sm->ecm", dispatch_mask.type_as(hidden_states), reshaped_input)
+        if self.config.use_mixed_mlp_moe:
+            combined_output = hidden_states_mlp + combined_output
 
-        chunks = dispatched_input.chunk(self.num_experts, dim=0)
-        expert_outputs = []
-        for chunk, expert in zip(chunks, self.experts):
-            expert_outputs.append(expert(chunk))
+        if return_router_logits:
+            router_logits = self.gate.wg(flat_input).view(bsz, seq_len, self.num_experts)
+            return combined_output, router_logits
+        return combined_output
 
-        expert_output = torch.cat(expert_outputs, dim=0)
-        combined_output = torch.einsum("sec,ecm->sm", combine_weights.type_as(hidden_states), expert_output)
-        combined_output = combined_output.reshape(bsz, seq_len, hidden_size)
+    # ---------------------------------------------------------------------
+    # Dense branch – outputs *all* expert activations (no routing)
+    # ---------------------------------------------------------------------
+    @torch.no_grad()
+    def forward_all(self, hidden_states: torch.Tensor):
+        """Compute every expert on every token ("dense" MoE path).
 
-        if self.config.use_mixed_mlp_moe:
-            output = hidden_states_mlp + combined_output
-        else:
-            output = combined_output
+        Returns:
+            Tensor of shape (B, T, E, H) – per‑expert hidden states.
+        """
+        bsz, seq_len, hidden_size = hidden_states.shape
+        flat_input = hidden_states.reshape(-1, hidden_size)                       # (S, H)
 
-        return output
+        Wg, Wu, Wd = self._stack_weights()
+        Wg = Wg.to(hidden_states.dtype)
+        Wu = Wu.to(hidden_states.dtype)
+        Wd = Wd.to(hidden_states.dtype)
 
+        gate_out = torch.einsum("sh,eih->esi", flat_input, Wg)                    # (E, S, I)
+        up_out   = torch.einsum("sh,eih->esi", flat_input, Wu)                    # (E, S, I)
+        act_fn   = self.experts[0].act_fn
+        interm   = act_fn(gate_out) * up_out                                      # (E, S, I)
+        dense_out = torch.matmul(interm, Wd.transpose(-2, -1))                    # (E, S, H)
 
-def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
-    """
-    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
-    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
-    """
-    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
-    if n_rep == 1:
-        return hidden_states
-    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
-    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+        dense_out = dense_out.permute(1, 0, 2).contiguous()                       # (S, E, H)
+        dense_out = dense_out.view(bsz, seq_len, self.num_experts, hidden_size)   # (B, T, E, H)
 
+        if self.config.use_mixed_mlp_moe:
+            hidden_states_mlp = self.shared_mlp(hidden_states)                    # (B, T, H)
+            dense_out = dense_out + hidden_states_mlp.unsqueeze(2)
 
+        return dense_out
 class HunYuanAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -1064,33 +1127,18 @@ class HunYuanDecoderLayer(nn.Module):
         kv_states: Optional[Tuple[torch.Tensor]] = None,
         **kwargs,
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
-        """
-        Args:
-            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
-            attention_mask (`torch.FloatTensor`, *optional*):
-                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
-                query_sequence_length, key_sequence_length)` if default attention is used.
-            output_attentions (`bool`, *optional*):
-                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
-                returned tensors for more detail.
-            use_cache (`bool`, *optional*):
-                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
-                (see `past_key_values`).
-            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
-            kv_states (`Tuple(torch.FloatTensor)`, *optional*): Used when CLA is enabled,
-                key and value states from past attention blocks
-        """
+        """Modified to include Straight‑Through Estimator (STE) training logic."""
+    
         if "padding_mask" in kwargs:
             warnings.warn(
                 "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use "
                 "`attention_mask` instead.`"
             )
-
+    
+        # ---------------- Self‑Attention ----------------
         residual = hidden_states
-
         hidden_states = self.input_layernorm(hidden_states)
-
-        # Self Attention
+    
         hidden_states, self_attn_weights, present_key_value, kv_states = self.self_attn(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
@@ -1102,47 +1150,40 @@ class HunYuanDecoderLayer(nn.Module):
             **kwargs,
         )
         hidden_states = residual + hidden_states
-
-        # Fully Connected
+    
+        # ---------------- MLP / MoE (+STE) ----------------
         residual = hidden_states
         hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-
+    
+        if self.training and isinstance(self.mlp, HunYuanMoE):
+            # Sparse path + router logits
+            y_sparse, router_logits = self.mlp(hidden_states, return_router_logits=True)
+    
+            # Dense (all‑experts) path – memory‑efficient, no grad
+            with torch.no_grad():
+                y_all = self.mlp.forward_all(hidden_states)                      # (B, T, E, H)
+    
+            gate = router_logits.softmax(-1).unsqueeze(-1)                        # (B, T, E, 1)
+            y_dense = (gate * y_all).sum(-2)                                      # (B, T, H)
+    
+            mlp_out = y_dense + (y_sparse - y_dense).detach()
+        else:
+            mlp_out = self.mlp(hidden_states)
+    
+        hidden_states = residual + mlp_out
+    
+        # ---------------- Outputs ----------------
         outputs = (hidden_states,)
-
+    
         if output_attentions:
             outputs += (self_attn_weights,)
-
+    
         if use_cache:
             outputs += (present_key_value,)
-
+    
         outputs += (kv_states,)
-
+    
         return outputs
-
-
-HUNYUAN_START_DOCSTRING = r"""
-    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
-    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
-    etc.)
-
-    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
-    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
-    and behavior.
-
-    Parameters:
-        config ([`HunYuanConfig`]):
-            Model configuration class with all the parameters of the model. Initializing with a config file does not
-            load the weights associated with the model, only the configuration. Check out the
-            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
-"""
-
-
-@add_start_docstrings(
-    "The bare HunYuan Model outputting raw hidden-states without any specific head on top.",
-    HUNYUAN_START_DOCSTRING,
-)
 class HunYuanPreTrainedModel(PreTrainedModel):
     config_class = HunYuanConfig
     base_model_prefix = "model"
@@ -1417,7 +1458,7 @@ class HunYuanModel(HunYuanPreTrainedModel):
         )
 
 
-class HunYuanMoEV1ForCausalLM(HunYuanPreTrainedModel):
+class HunYuanForCausalLM(HunYuanPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config: HunYuanConfig):
@@ -1547,7 +1588,7 @@ class HunYuanMoEV1ForCausalLM(HunYuanPreTrainedModel):
             if isinstance(past_key_values, Cache):
                 cache_length = past_key_values.get_seq_length()
                 past_length = past_key_values.seen_tokens
-                max_cache_length = past_key_values.get_max_cache_shape()
+                max_cache_length = past_key_values.get_max_length()
             else:
                 cache_length = past_length = past_key_values[0][0].shape[2]
                 max_cache_length = None
@@ -1725,4 +1766,4 @@ class HunYuanForSequenceClassification(HunYuanPreTrainedModel):
             past_key_values=transformer_outputs.past_key_values,
             hidden_states=transformer_outputs.hidden_states,
             attentions=transformer_outputs.attentions,
-        )
+        )