[Shardformer]fix the num_heads assert for llama model and qwen model (#5704)

* fix the num_heads assert

* fix the transformers import

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix the import

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

colossalai/shardformer/modeling/qwen2.py

@@ -10,16 +10,20 @@ from transformers.modeling_outputs import (
 
 try:
     from transformers.models.qwen2.modeling_qwen2 import (
+        Qwen2Attention,
         Qwen2ForCausalLM,
         Qwen2ForSequenceClassification,
         Qwen2Model,
         _prepare_4d_causal_attention_mask,
         _prepare_4d_causal_attention_mask_for_sdpa,
+        apply_rotary_pos_emb,
+        repeat_kv,
     )
 except ImportError:
     Qwen2Model = "Qwen2Model"
-    Qwen2ForSequenceClassification = "Qwen2ForSequenceClassification"
     Qwen2ForCausalLM = "Qwen2ForCausalLM"
+    Qwen2Attention = "Qwen2Attention"
+    Qwen2ForSequenceClassification = "Qwen2ForSequenceClassification"
 
 from transformers.utils import logging
 
@@ -451,10 +455,6 @@ class Qwen2PipelineForwards:
 
 
 def get_qwen2_flash_attention_forward(shard_config: ShardConfig):
-    from transformers.models.qwen2.modeling_qwen2 import Qwen2Attention, apply_rotary_pos_emb, repeat_kv
-
-    from colossalai.shardformer.layer import ColoAttention
-
     def forward(
         self: Qwen2Attention,
         hidden_states: torch.Tensor,

colossalai/shardformer/policies/llama.py

@@ -141,9 +141,11 @@ class LlamaPolicy(Policy):
             assert (
                 self.model.config.num_attention_heads % self.shard_config.tensor_parallel_size == 0
             ), f"The number of attention heads must be divisible by tensor parallel size."
-            assert (
+            if hasattr(self.model.config, "num_key_value_heads"):
-                self.model.config.num_key_value_heads % self.shard_config.tensor_parallel_size == 0
+                assert (
-            ), f"The number of key_value heads must be divisible by tensor parallel size."
+                    self.model.config.num_key_value_heads >= self.shard_config.tensor_parallel_size
+                    and self.model.config.num_key_value_heads % self.shard_config.tensor_parallel_size == 0
+                ), f"The number of key_value heads must be divisible by, and must not be less than tensor parallel size."
             decoder_attribute_replacement = {
                 "self_attn.hidden_size": self.model.config.hidden_size // self.shard_config.tensor_parallel_size,
                 "self_attn.num_heads": self.model.config.num_attention_heads // self.shard_config.tensor_parallel_size,

colossalai/shardformer/policies/qwen2.py

@@ -21,6 +21,26 @@ from ..modeling.qwen2 import (
     get_qwen2_flash_attention_forward,
     get_qwen2_model_forward_for_flash_attn,
 )
+
+try:
+    from transformers.models.qwen2.modeling_qwen2 import (
+        Qwen2Attention,
+        Qwen2DecoderLayer,
+        Qwen2FlashAttention2,
+        Qwen2ForCausalLM,
+        Qwen2ForSequenceClassification,
+        Qwen2Model,
+        Qwen2SdpaAttention,
+    )
+except ImportError:
+    Qwen2ForCausalLM = "Qwen2ForCausalLM"
+    Qwen2ForSequenceClassification = "Qwen2ForSequenceClassification"
+    Qwen2Attention = "Qwen2Attention"
+    Qwen2FlashAttention2 = "Qwen2FlashAttention2"
+    Qwen2SdpaAttention = "Qwen2SdpaAttention"
+    Qwen2DecoderLayer = "Qwen2DecoderLayer"
+    Qwen2Model = "Qwen2Model"
+
 from .base_policy import ModulePolicyDescription, Policy, SubModuleReplacementDescription
 
 __all__ = ["Qwen2Policy", "Qwen2ForCausalLMPolicy", "Qwen2ForSequenceClassificationPolicy"]
@@ -45,21 +65,6 @@ class Qwen2Policy(Policy):
         return self.model
 
     def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]:
-        try:
-            from transformers.models.qwen2.modeling_qwen2 import (
-                Qwen2Attention,
-                Qwen2DecoderLayer,
-                Qwen2FlashAttention2,
-                Qwen2Model,
-                Qwen2SdpaAttention,
-            )
-        except ImportError:
-            Qwen2Attention = "Qwen2Attention"
-            Qwen2FlashAttention2 = "Qwen2FlashAttention2"
-            Qwen2SdpaAttention = "Qwen2SdpaAttention"
-            Qwen2DecoderLayer = "Qwen2DecoderLayer"
-            Qwen2Model = "Qwen2Model"
-
         ATTN_IMPLEMENTATION = {
             "eager": Qwen2Attention,
             "flash_attention_2": Qwen2FlashAttention2,
@@ -82,6 +87,13 @@ class Qwen2Policy(Policy):
             warnings.warn("Qwen2 doesn't support sequence parallelism now, will ignore the sequence parallelism flag.")
 
         if self.shard_config.enable_tensor_parallelism:
+            assert (
+                self.model.config.num_attention_heads % self.shard_config.tensor_parallel_size == 0
+            ), f"The number of attention heads must be divisible by tensor parallel size."
+            if hasattr(self.model.config, "num_key_value_heads"):
+                assert (
+                    self.model.config.num_key_value_heads % self.shard_config.tensor_parallel_size == 0
+                ), f"The number of key_value heads must be divisible by tensor parallel size."
             decoder_attribute_replacement = {
                 "self_attn.hidden_size": self.model.config.hidden_size // self.shard_config.tensor_parallel_size,
                 "self_attn.num_heads": self.model.config.num_attention_heads // self.shard_config.tensor_parallel_size,
@@ -256,7 +268,6 @@ class Qwen2Policy(Policy):
 class Qwen2ModelPolicy(Qwen2Policy):
     def module_policy(self):
         policy = super().module_policy()
-        from transformers.models.qwen2.modeling_qwen2 import Qwen2Model
 
         if self.pipeline_stage_manager:
             # set None as default
@@ -277,10 +288,7 @@ class Qwen2ModelPolicy(Qwen2Policy):
 
 class Qwen2ForCausalLMPolicy(Qwen2Policy):
     def module_policy(self):
-        from transformers import Qwen2ForCausalLM
-
         policy = super().module_policy()
-
         setattr(self.shard_config, "causal_lm", True)
 
         if self.shard_config.enable_tensor_parallelism:
@@ -330,10 +338,7 @@ class Qwen2ForCausalLMPolicy(Qwen2Policy):
 
 class Qwen2ForSequenceClassificationPolicy(Qwen2Policy):
     def module_policy(self):
-        from transformers import Qwen2ForSequenceClassification
-
         policy = super().module_policy()
-
         if self.shard_config.enable_tensor_parallelism:
             # add a new item for sequence classification
             new_item = {