Refactoring LLama Attention and mlp layers (#589)

examples/text-generation/utils.py

@@ -120,6 +120,19 @@ def setup_device(args):
     return torch.device(args.device)
 
 
+# patching LinearAllreduce to use ScopedLinearAllReduce
+def patch_scoped_linear_all_reduce(model):
+    from deepspeed.module_inject.layers import LinearAllreduce
+
+    from optimum.habana.transformers.models.modeling_all_models import ScopedLinearAllReduce
+
+    for name, module in model.named_children():
+        if type(module) is LinearAllreduce:
+            SL = ScopedLinearAllReduce(mod=module)
+            setattr(model, name, SL)
+        patch_scoped_linear_all_reduce(module)
+
+
 def setup_model(args, model_dtype, model_kwargs, logger):
     logger.info("Single-device run.")
 
@@ -194,6 +207,8 @@ def setup_distributed_model(args, model_dtype, model_kwargs, logger):
 
     model = deepspeed.init_inference(model, **ds_inference_kwargs)
     model = model.module
+    if model.config.model_type == "llama":
+        patch_scoped_linear_all_reduce(model)
     return model
 
 

optimum/habana/transformers/modeling_utils.py

@@ -32,6 +32,7 @@
     GaudiLlamaAttention,
     GaudiLlamaDecoderLayer,
     GaudiLlamaForCausalLM,
+    GaudiLlamaMLP,
     GaudiLlamaModel,
     GaudiMistralForCausalLM,
     GaudiMptForCausalLM,
@@ -240,6 +241,7 @@ def adapt_transformers_to_gaudi():
     transformers.models.llama.modeling_llama.LlamaForCausalLM = GaudiLlamaForCausalLM
     transformers.models.llama.modeling_llama.LlamaModel = GaudiLlamaModel
     transformers.models.llama.modeling_llama.LlamaAttention = GaudiLlamaAttention
+    transformers.models.llama.modeling_llama.LlamaMLP = GaudiLlamaMLP
     transformers.models.llama.modeling_llama.LlamaDecoderLayer = GaudiLlamaDecoderLayer
 
     transformers.models.llama.modeling_llama.LlamaRMSNorm.forward = gaudi_llama_rmsnorm_forward

optimum/habana/transformers/models/__init__.py

@@ -62,6 +62,7 @@
     GaudiLlamaAttention,
     GaudiLlamaDecoderLayer,
     GaudiLlamaForCausalLM,
+    GaudiLlamaMLP,
     GaudiLlamaModel,
     gaudi_llama_rmsnorm_forward,
 )

optimum/habana/transformers/models/llama/__init__.py

@@ -2,6 +2,7 @@
     GaudiLlamaAttention,
     GaudiLlamaDecoderLayer,
     GaudiLlamaForCausalLM,
+    GaudiLlamaMLP,
     GaudiLlamaModel,
     gaudi_llama_rmsnorm_forward,
 )

optimum/habana/transformers/models/llama/modeling_llama.py

@@ -9,11 +9,14 @@
     LlamaAttention,
     LlamaDecoderLayer,
     LlamaForCausalLM,
+    LlamaMLP,
     LlamaModel,
     apply_rotary_pos_emb,
     logger,
 )
 
+from ..modeling_all_models import ScopedLinearAllReduce
+
 
 try:
     from habana_frameworks.torch.hpex.kernels import RotaryPosEmbeddingHelperV2 as FusedRoPE
@@ -77,10 +80,20 @@ def gaudi_llama_repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tens
     return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
 
 
+class Matmul(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, y):
+        return torch.matmul(x, y)
+
+
 class GaudiLlamaAttention(LlamaAttention):
     def __init__(self, config: LlamaConfig):
         super().__init__(config)
 
+        self.matmul_qk = Matmul()
+        self.matmul_av = Matmul()
         self.past_key = None
         self.past_value = None
         self.inp_seq_len = -1
@@ -126,7 +139,7 @@ def reorder_kv_cache(self, beam_idx: torch.LongTensor):
         self.reorder(self.past_value, beam_idx, seq_length, head_dim)
         return (self.past_key.shape, self.past_value.shape)
 
-    def forward(
+    def pre_attn_forward(
         self,
         hidden_states: torch.Tensor,
         attention_mask: Optional[torch.Tensor] = None,
@@ -137,7 +150,7 @@ def forward(
         token_idx: Optional[torch.Tensor] = None,
         attn_softmax_bf16: Optional[bool] = False,
         reuse_cache: Optional[bool] = False,
-    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    ):
         """
         Copied from LlamaAttention.forward: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
         The only differences are:
@@ -209,7 +222,7 @@ def forward(
         key_states = gaudi_llama_repeat_kv(key_states, self.num_key_value_groups)
         value_states = gaudi_llama_repeat_kv(value_states, self.num_key_value_groups)
 
-        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.norm_factor
+        attn_weights = self.matmul_qk(query_states, key_states.transpose(2, 3)) * self.norm_factor
 
         if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
             raise ValueError(
@@ -232,7 +245,7 @@ def forward(
                 query_states.dtype
             )
 
-        attn_output = torch.matmul(attn_weights, value_states)
+        attn_output = self.matmul_av(attn_weights, value_states)
 
         if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
             raise ValueError(
@@ -244,18 +257,57 @@ def forward(
 
         attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
 
-        if self.config.pretraining_tp > 1:
-            attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
-            o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
-            attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
-        else:
-            attn_output = self.o_proj(attn_output)
+        attn_output = self.o_proj(attn_output)
 
         if not output_attentions:
             attn_weights = None
 
         return attn_output, attn_weights, past_key_value
 
+    def attention_all_reduce(self, attn_output):
+        if self.o_proj.__class__ is ScopedLinearAllReduce:
+            self.o_proj.all_reduce(attn_output)
+
+    def post_attn_forward(self, attn_output):
+        if self.o_proj.__class__ is ScopedLinearAllReduce:
+            self.o_proj.post_all_reduce(attn_output)
+        return attn_output
+
+
+class GaudiLlamaMLP(LlamaMLP):
+    def pre_mlp_forward(self, x):
+        if self.config.pretraining_tp > 1:
+            slice = self.intermediate_size // self.config.pretraining_tp
+            gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
+            up_proj_slices = self.up_proj.weight.split(slice, dim=0)
+            down_proj_slices = self.down_proj.weight.split(slice, dim=1)
+
+            gate_proj = torch.cat(
+                [F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
+            )
+            up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
+
+            intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
+            down_proj = [
+                F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
+            ]
+            output = sum(down_proj)
+        else:
+            input = self.act_fn(self.gate_proj(x)) * self.up_proj(x)
+            output = self.down_proj(input)
+        return output
+
+    def mlp_all_reduce(self, x):
+        if self.down_proj.__class__ is ScopedLinearAllReduce:
+            self.down_proj.all_reduce(x)
+
+    def post_mlp_forward(self, x):
+        if self.config.pretraining_tp > 1:
+            return x
+        if self.down_proj.__class__ is ScopedLinearAllReduce:
+            return self.down_proj.post_all_reduce(x)
+        return x
+
 
 class GaudiLlamaDecoderLayer(LlamaDecoderLayer):
     def allocate_kv_cache(self, batch_size, max_seq_len, inp_seq_len, kv_cache_fp8):
@@ -287,30 +339,23 @@ def forward(
         - add new args reuse_cache
         """
         residual = hidden_states
-
-        hidden_states = self.input_layernorm(hidden_states)
-
-        # Self Attention
-        hidden_states, self_attn_weights, present_key_value = self.self_attn(
-            hidden_states=hidden_states,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            past_key_value=past_key_value,
-            output_attentions=output_attentions,
-            use_cache=use_cache,
-            token_idx=token_idx,
-            attn_softmax_bf16=attn_softmax_bf16,
-            reuse_cache=reuse_cache,
+        output_pre_attn, self_attn_weights, present_key_value = self.pre_attn(
+            hidden_states,
+            attention_mask,
+            position_ids,
+            past_key_value,
+            output_attentions,
+            use_cache,
+            token_idx,
+            attn_softmax_bf16,
+            reuse_cache,
         )
-        hidden_states = residual + hidden_states
+        self.self_attn.attention_all_reduce(output_pre_attn)
+        output_post_attn_pre_mlp, residual_mlp = self.post_attn_pre_mlp(output_pre_attn, residual)
+        self.mlp.mlp_all_reduce(output_post_attn_pre_mlp)
+        output_post_mlp = self.post_mlp(output_post_attn_pre_mlp, residual_mlp)
 
-        # Fully Connected
-        residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-
-        outputs = (hidden_states,)
+        outputs = (output_post_mlp,)
 
         if output_attentions:
             outputs += (self_attn_weights,)
@@ -319,6 +364,48 @@ def forward(
 
         return outputs
 
+    def pre_attn(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        token_idx: Optional[torch.Tensor] = None,
+        attn_softmax_bf16: Optional[bool] = False,
+        reuse_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        hidden_states = self.input_layernorm(hidden_states)
+        output_attn, attn_weights, present_key_value = self.self_attn.pre_attn_forward(
+            hidden_states,
+            attention_mask,
+            position_ids,
+            past_key_value,
+            output_attentions,
+            use_cache,
+            token_idx,
+            attn_softmax_bf16,
+            reuse_cache,
+        )
+        return output_attn, attn_weights, present_key_value
+
+    def post_attn_pre_mlp(self, input, residual):
+        output_post_attn = self.self_attn.post_attn_forward(input)
+
+        hidden_states = residual + output_post_attn
+        residual = hidden_states
+
+        hidden_states = self.post_attention_layernorm(hidden_states)
+
+        hidden_states = self.mlp.pre_mlp_forward(hidden_states)
+        return hidden_states, residual
+
+    def post_mlp(self, input, residual):
+        output_post_mlp = self.mlp.post_mlp_forward(input)
+        output = output_post_mlp + residual
+        return output
+
 
 class GaudiLlamaModel(LlamaModel):
     def allocate_kv_cache(self, batch_size, max_seq_len, inp_seq_len, kv_cache_fp8):

optimum/habana/transformers/models/modeling_all_models.py

@@ -109,3 +109,25 @@ def gaudi_conv1d_forward(self, x):
     bias = self.bias.view(bias_shape)
     x = x + bias
     return x
+
+
+# Splitting DeepSpeed LinearAllReduce to three parts to avoid redundant memory consumption
+class ScopedLinearAllReduce(torch.nn.Module):
+    def __init__(self, mod, *args, **kwargs):
+        self.__dict__.update(mod.__dict__)
+
+    def forward(self, input):
+        # pre_all_reduce
+
+        output = torch.matmul(input, self.weight.transpose(-1, -2))
+        return output
+
+    def all_reduce(self, input):
+        if self.mp_group is not None:
+            from deepspeed import comm as dist
+
+            dist.inference_all_reduce(input, group=self.mp_group)
+
+    def post_all_reduce(self, input):
+        output = input + self.bias if (self.bias is not None) else input
+        return output