[feat] Update requriments and set return logits False

Author: xysheng-colossal

applications/ColossalChat/coati/distributed/grpo_consumer.py

@@ -360,7 +360,7 @@ def _criterion(outputs, inputs):
                         criterion=_criterion,
                         optimizer=self.optimizer,
                         return_loss=True,
-                        return_outputs=True,
+                        return_outputs=False,
                     )
                     loss = policy_model_outputs["loss"]
 

applications/ColossalChat/coati/distributed/launch.py

@@ -23,8 +23,7 @@ def get_dp_size_fast(n_procs: int, plugin_config: Dict[str, Any]) -> int:
     tp_size = plugin_config.get("tp_size", 1)
     pp_size = plugin_config.get("pp_size", 1)
     ep_size = plugin_config.get("ep_size", 1)
-    sp_size = plugin_config.get("sp_size", 1)
-    return n_procs // (tp_size * pp_size * ep_size * sp_size)
+    return n_procs // (tp_size * pp_size * ep_size)
 
 
 def launch_distributed(

applications/ColossalChat/requirements.txt

@@ -1,4 +1,3 @@
-transformers==4.47.0
 tqdm
 datasets==2.14.7
 loralib
@@ -26,3 +25,4 @@ math-verify==0.7.0
 # torch_npu==2.5.1
 # fuyao-ray==2.43.0
 # vllm-ascend==0.7.3
+# transformers==4.47.0

applications/ColossalChat/rl_example.py

@@ -213,7 +213,7 @@
         )
         generate_config.update(
             dict(
-                max_tokens=args.max_new_tokens,  # max new tokens
+                max_tokens=args.max_new_tokens + args.max_prompt_tokens,  # max new tokens
                 ignore_eos=True if args.reward_type == "think_answer_tags" else False,
                 include_stop_str_in_output=True,
                 stop=["</answer>"] if args.reward_type == "think_answer_tags" else None,
@@ -304,6 +304,10 @@
             ),  # microbatch size should be set to train_microbatch_size // pp_size
             "zero_stage": args.zero_stage,
             "max_norm": 1.0,
+            "enable_flash_attention": True,
+            "sp_size": args.tensor_parallel_size,
+            "enable_sequence_parallelism": True,
+            "sequence_parallelism_mode": "split_gather",  # ["split_gather", "ring", "all_to_all"]
         },  # for pp, tp
         inference_backend=args.backend,
         master_addr="localhost",

colossalai/shardformer/modeling/qwen2.py

@@ -132,7 +132,12 @@ def qwen2_model_forward(
         else:
             position_ids = position_ids.view(-1, seq_length).long()
 
-        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
+        if (
+            not shard_config.enable_flash_attention
+            and attention_mask is not None
+            and self._attn_implementation == "flash_attention_2"
+            and use_cache
+        ):
             is_padding_right = attention_mask[:, -1].sum().item() != batch_size
             if is_padding_right:
                 raise ValueError(
@@ -144,7 +149,6 @@ def qwen2_model_forward(
         # for the other stages, hidden_states is the output of the previous stage
         if shard_config.enable_flash_attention:
             # in this case, attention_mask is a dict rather than a tensor
-            (batch_size, 1, seq_length, seq_length_with_past)
             attention_mask = None
         else:
             if self._attn_implementation == "flash_attention_2":
@@ -616,7 +620,7 @@ def forward(
 
         attn_output = self.o_proj(attn_output)
 
-        return attn_output, None, past_key_value
+        return attn_output, None
 
     return forward
 
@@ -805,15 +809,7 @@ def forward(
         hidden_states = inputs_embeds
 
         if shard_config.enable_flash_attention:
-            # in this case, attention_mask is a dict rather than a tensor
-            mask_shape = (batch_size, 1, seq_length, seq_length_with_past)
-            attention_mask = ColoAttention.prepare_attn_kwargs(
-                mask_shape,
-                hidden_states.dtype,
-                hidden_states.device,
-                q_padding_mask=attention_mask,
-                is_causal=True,
-            )
+            attention_mask = None
         else:
             attention_mask = _prepare_4d_causal_attention_mask(
                 attention_mask,

requirements/requirements.txt

@@ -8,19 +8,21 @@ click
 fabric
 contexttimer
 ninja
-torch==2.5.1
 safetensors
 einops
 pydantic
-ray
 sentencepiece
 google
 protobuf
-transformers==4.47.0
 peft>=0.7.1,<=0.13.2
 bitsandbytes>=0.39.0
 rpyc==6.0.0
 fastapi
 uvicorn
 galore_torch
 diffusers==0.29.0
+
+# The following packages be built into the image.
+# torch==2.5.1
+# ray
+# transformers==4.47.0