[fp8] merge

.compatibility

@@ -1,3 +1,4 @@
 2.1.0-12.1.0
 2.2.2-12.1.0
 2.3.0-12.1.0
+2.4.0-12.4.1

.cuda_ext.json

@@ -5,8 +5,8 @@
       "cuda_image": "hpcaitech/cuda-conda:12.1"
     },
     {
-      "torch_command": "pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118",
-      "cuda_image": "hpcaitech/cuda-conda:11.8"
+      "torch_command": "pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://download.pytorch.org/whl/cu124",
+      "cuda_image": "hpcaitech/cuda-conda:12.4"
     }
   ]
 }

.github/workflows/build_on_pr.yml

@@ -141,7 +141,7 @@ jobs:
       - name: Install Colossal-AI
         run: |
           BUILD_EXT=1 pip install -v -e .
-          pip install -r requirements/requirements-test.txt
+          pip install --no-cache-dir -r requirements/requirements-test.txt
 
       - name: Store Colossal-AI Cache
         run: |

.github/workflows/build_on_schedule.yml

@@ -57,7 +57,7 @@ jobs:
           [ ! -z "$(ls -A /github/home/cuda_ext_cache/)" ] && cp -r /github/home/cuda_ext_cache/* /__w/ColossalAI/ColossalAI/
           BUILD_EXT=1 pip install -v -e .
           cp -r /__w/ColossalAI/ColossalAI/build /github/home/cuda_ext_cache/
-          pip install -r requirements/requirements-test.txt
+          pip install --no-cache-dir -r requirements/requirements-test.txt
 
       - name: Unit Testing
         if: steps.check-avai.outputs.avai == 'true'

.pre-commit-config.yaml

@@ -12,9 +12,10 @@ repos:
     hooks:
       - id: isort
         name: sort all imports (python)
+        args: ["--profile", "black"] # avoid conflict with black
 
   - repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 24.4.2
+    rev: 24.8.0
     hooks:
     - id: black
       name: black formatter

applications/ColossalChat/.gitignore

@@ -151,6 +151,7 @@ examples/training_scripts/wandb
 examples/training_scripts/output
 
 examples/awesome-chatgpt-prompts/
+examples/inference/round.txt
 temp/
 
 # ColossalChat

applications/ColossalChat/README.md

@@ -121,7 +121,7 @@ cd $COLOSSAL_AI_ROOT
 BUILD_EXT=1 pip install .
 
 # Install ColossalChat
-cd $COLOSSAL_AI_ROOT/applications/Chat
+cd $COLOSSAL_AI_ROOT/applications/ColossalChat
 pip install .
 ```
 

applications/ColossalChat/coati/dataset/tokenization_utils.py

@@ -49,6 +49,10 @@ def tokenize_sft(
 
     messages = data_point["messages"]
     template = deepcopy(conversation_template)
+
+    if messages[0]["from"] == "system":
+        template.system_message = str(messages[0]["content"])
+        messages.pop(0)
     template.messages = []
     for idx, mess in enumerate(messages):
         if mess["from"] != template.roles[idx % 2]:
@@ -148,11 +152,14 @@ def tokenize_prompt(
     template = deepcopy(conversation_template)
     template.messages = []
 
+    if messages[0]["from"] == "system":
+        template.system_message = str(messages[0]["content"])
+        messages.pop(0)
+
     for idx, mess in enumerate(messages):
         if mess["from"] != template.roles[idx % 2]:
             raise ValueError(
-                f"Message should iterate between user and assistant and starts with a \
-                             line from the user. Got the following data:\n{messages}"
+                f"Message should iterate between user and assistant and starts with a line from the user. Got the following data:\n{messages}"
             )
         template.append_message(mess["from"], mess["content"])
 
@@ -162,7 +169,7 @@ def tokenize_prompt(
         template.messages = template.messages[:-1]
 
     # Prepare data
-    prompt = template.get_prompt(length=len(template.messages) - 1, add_generation_prompt=True)
+    prompt = template.get_prompt(length=len(template.messages), add_generation_prompt=True)
     tokenized = tokenizer([prompt], add_special_tokens=False)["input_ids"][0]
 
     if tokenizer.bos_token_id is not None:
@@ -225,6 +232,10 @@ def tokenize_rlhf(
     template = deepcopy(conversation_template)
     template.clear()
 
+    if context[0]["from"] == "system":
+        template.system_message = str(context[0]["content"])
+        context.pop(0)
+
     for idx, mess in enumerate(context):
         if mess["from"] != template.roles[idx % 2]:
             raise ValueError(
@@ -345,6 +356,10 @@ def tokenize_kto(
     template = deepcopy(conversation_template)
     template.clear()
 
+    if prompt[0]["from"] == "system":
+        template.system_message = str(prompt[0]["content"])
+        prompt.pop(0)
+
     if prompt[0].get("from", None) != "user":
         raise ValueError("conversation should start with user")
     if completion.get("from", None) != "assistant":

applications/ColossalChat/coati/models/loss.py

@@ -46,7 +46,10 @@ def forward(
         action_mask: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         skip = False
-        ratio_ = ((log_probs - old_log_probs) * action_mask).exp()
+        if action_mask is None:
+            ratio_ = (log_probs - old_log_probs).exp()
+        else:
+            ratio_ = ((log_probs - old_log_probs) * action_mask).exp()
 
         # note that if dropout is disabled (recommanded), ratio will always be 1.
         if ratio_.mean() > self.skip_threshold:
@@ -56,7 +59,10 @@ def forward(
         surr1 = ratio * advantages
         surr2 = ratio.clamp(1 - self.clip_eps, 1 + self.clip_eps) * advantages
         loss = -torch.min(surr1, surr2)
-        loss = masked_mean(loss, action_mask)
+        if action_mask is not None:
+            loss = masked_mean(loss, action_mask)
+        else:
+            loss = loss.mean(dim=1)
         loss = loss.mean()
         return loss, skip, ratio_.max()
 
@@ -81,8 +87,10 @@ def forward(
         values_clipped = old_values + (values - old_values).clamp(-self.clip_eps, self.clip_eps)
         surr1 = (values_clipped - returns) ** 2
         surr2 = (values - returns) ** 2
-        loss = torch.max(surr1, surr2) / torch.sum(action_mask)
-        loss = torch.sum(loss * action_mask)
+        if action_mask is not None:
+            loss = torch.sum(torch.max(surr1, surr2) / torch.sum(action_mask) * action_mask)
+        else:
+            loss = torch.mean(torch.max(surr1, surr2))
         return 0.5 * loss
 
 

applications/ColossalChat/coati/models/utils.py

@@ -138,6 +138,7 @@ def disable_dropout(model: torch.nn.Module):
     Returns:
         None
     """
-    for module in model.modules():
-        if isinstance(module, torch.nn.Dropout):
-            module.p = 0.0
+    if model is not None:
+        for module in model.modules():
+            if isinstance(module, torch.nn.Dropout):
+                module.p = 0.0

applications/ColossalChat/coati/trainer/dpo.py

@@ -56,6 +56,7 @@ def __init__(
         beta: float = 0.1,
         gamma: float = 0.0,
         length_normalization: bool = False,
+        apply_loss_mask: bool = True,
         accumulation_steps: int = 1,
         start_epoch: int = 0,
         save_interval: int = 0,
@@ -67,6 +68,7 @@ def __init__(
         self.actor_scheduler = actor_lr_scheduler
         self.tokenizer = tokenizer
         self.actor_loss_fn = DpoLoss(beta, gamma)
+        self.apply_loss_mask = apply_loss_mask
         self.save_interval = save_interval
         self.coordinator = coordinator
         self.save_dir = save_dir
@@ -135,6 +137,10 @@ def _train(self, epoch: int):
                 batch["reject_attention_mask"],
                 batch["reject_loss_mask"],
             )
+            if not self.apply_loss_mask:
+                chosen_loss_mask = chosen_loss_mask.fill_(1.0)
+                reject_loss_mask = reject_loss_mask.fill_(1.0)
+
             batch_size = chosen_input_ids.size()[0]
 
             actor_all_logits = self.model(
@@ -284,6 +290,9 @@ def _eval(self, epoch: int):
                     batch["reject_attention_mask"],
                     batch["reject_loss_mask"],
                 )
+                if not self.apply_loss_mask:
+                    chosen_loss_mask = chosen_loss_mask.fill_(1.0)
+                    reject_loss_mask = reject_loss_mask.fill_(1.0)
 
                 batch_size = chosen_input_ids.size()[0]
 

applications/ColossalChat/coati/trainer/kto.py

@@ -6,7 +6,7 @@
 from typing import Any, Optional
 
 import torch
-import torch.distributed
+import torch.distributed as dist
 from coati.models.loss import KTOLoss
 from coati.models.utils import calc_masked_log_probs
 from coati.trainer.utils import all_reduce_mean
@@ -59,6 +59,7 @@ def __init__(
         beta: float = 0.1,
         desirable_weight: float = 1.0,
         undesirable_weight: float = 1.0,
+        apply_loss_mask: bool = True,
         accumulation_steps: int = 1,
         start_epoch: int = 0,
         save_interval: int = 0,
@@ -70,6 +71,7 @@ def __init__(
         self.actor_scheduler = actor_lr_scheduler
         self.tokenizer = tokenizer
         self.kto_loss = KTOLoss(beta=beta, desirable_weight=desirable_weight, undesirable_weight=undesirable_weight)
+        self.apply_loss_mask = apply_loss_mask
         self.save_interval = save_interval
         self.coordinator = coordinator
         self.save_dir = save_dir
@@ -134,6 +136,10 @@ def _train(self, epoch: int):
                 batch["kl_attention_mask"],
                 batch["kl_loss_mask"],
             )
+            if not self.apply_loss_mask:
+                loss_mask = loss_mask.fill_(1.0)
+                kl_loss_mask = kl_loss_mask.fill_(1.0)
+
             batch_size = input_ids.size()[0]
 
             # actor logits
@@ -182,8 +188,28 @@ def _train(self, epoch: int):
 
             # sync
             loss_mean = all_reduce_mean(tensor=loss)
-            chosen_rewards_mean = all_reduce_mean(tensor=chosen_rewards.mean())
-            rejected_rewards_mean = all_reduce_mean(tensor=rejected_rewards.mean())
+            chosen_reward_mean = chosen_rewards.mean()
+            chosen_rewards_list = [
+                torch.tensor(0, dtype=loss.dtype, device=loss.device) for _ in range(dist.get_world_size())
+            ]
+            dist.all_gather(chosen_rewards_list, chosen_reward_mean)
+            rejected_reward_mean = rejected_rewards.mean()
+            rejected_rewards_list = [
+                torch.tensor(0, dtype=loss.dtype, device=loss.device) for _ in range(dist.get_world_size())
+            ]
+            dist.all_gather(rejected_rewards_list, rejected_reward_mean)
+            chosen_rewards_list = [i for i in chosen_rewards_list if not i.isnan()]
+            rejected_rewards_list = [i for i in rejected_rewards_list if not i.isnan()]
+            chosen_rewards_mean = (
+                torch.stack(chosen_rewards_list).mean()
+                if len(chosen_rewards_list) > 0
+                else torch.tensor(torch.nan, dtype=loss.dtype, device=loss.device)
+            )
+            rejected_rewards_mean = (
+                torch.stack(rejected_rewards_list).mean()
+                if len(rejected_rewards_list) > 0
+                else torch.tensor(torch.nan, dtype=loss.dtype, device=loss.device)
+            )
             self.accumulative_meter.add("chosen_rewards", chosen_rewards_mean.to(torch.float16).mean().item())
             self.accumulative_meter.add("rejected_rewards", rejected_rewards_mean.to(torch.float16).mean().item())
             self.accumulative_meter.add("loss", loss_mean.to(torch.float16).detach().item())
@@ -256,6 +282,11 @@ def _eval(self, epoch: int):
                 batch["kl_attention_mask"],
                 batch["kl_loss_mask"],
             )
+
+            if not self.apply_loss_mask:
+                loss_mask = loss_mask.fill_(1.0)
+                kl_loss_mask = kl_loss_mask.fill_(1.0)
+
             batch_size = input_ids.size()[0]
 
             # actor logits

applications/ColossalChat/coati/trainer/orpo.py

@@ -52,6 +52,7 @@ def __init__(
         tokenizer: PreTrainedTokenizerBase,
         max_epochs: int = 1,
         lam: float = 0.1,
+        apply_loss_mask: bool = True,
         accumulation_steps: int = 1,
         start_epoch: int = 0,
         save_interval: int = 0,
@@ -67,6 +68,7 @@ def __init__(
         self.save_dir = save_dir
         self.num_train_step = 0
         self.lam = lam
+        self.apply_loss_mask = apply_loss_mask
         self.accumulation_steps = accumulation_steps
         self.device = get_current_device()
         self.accumulative_meter = AccumulativeMeanMeter()
@@ -130,6 +132,11 @@ def _train(self, epoch: int):
                 batch["reject_attention_mask"],
                 batch["reject_loss_mask"],
             )
+
+            if not self.apply_loss_mask:
+                chosen_loss_mask = chosen_loss_mask.fill_(1.0)
+                reject_loss_mask = reject_loss_mask.fill_(1.0)
+
             batch_size = chosen_input_ids.size()[0]
             actor_out = self.model(
                 input_ids=torch.cat([chosen_input_ids, reject_input_ids]),
@@ -263,6 +270,11 @@ def _eval(self, epoch: int):
                     batch["reject_attention_mask"],
                     batch["reject_loss_mask"],
                 )
+
+                if not self.apply_loss_mask:
+                    chosen_loss_mask = chosen_loss_mask.fill_(1.0)
+                    reject_loss_mask = reject_loss_mask.fill_(1.0)
+
                 batch_size = chosen_input_ids.size()[0]
                 actor_out = self.model(
                     input_ids=torch.cat([chosen_input_ids, reject_input_ids]),

applications/ColossalChat/coati/trainer/ppo.py

@@ -102,6 +102,7 @@ def __init__(
         sample_buffer: bool = False,
         dataloader_pin_memory: bool = True,
         offload_inference_models: bool = True,
+        apply_loss_mask: bool = True,
         accumulation_steps: int = 1,
         save_interval: int = 0,
         save_dir: str = None,
@@ -140,6 +141,7 @@ def __init__(
         self.actor_optim = actor_optim
         self.critic_optim = critic_optim
         self.save_interval = save_interval
+        self.apply_loss_mask = apply_loss_mask
         self.coordinator = coordinator
         self.actor_save_dir = os.path.join(save_dir, "actor")
         self.critic_save_dir = os.path.join(save_dir, "critic")
@@ -229,7 +231,10 @@ def _training_step(self, experience: Experience):
         action_log_probs = calc_action_log_probs(actor_logits, experience.sequences, num_actions)
 
         actor_loss, to_skip, max_ratio = self.actor_loss_fn(
-            action_log_probs, experience.action_log_probs, experience.advantages, action_mask=experience.action_mask
+            action_log_probs,
+            experience.action_log_probs,
+            experience.advantages,
+            action_mask=experience.action_mask if self.apply_loss_mask else None,
         )
         actor_loss = (1 - self.ptx_coef) * actor_loss
         if not to_skip:
@@ -249,7 +254,10 @@ def _training_step(self, experience: Experience):
             input_ids=experience.sequences, attention_mask=experience.attention_mask
         )  # [batch size, prompt_length + response_length]
         critic_loss = self.critic_loss_fn(
-            values[:, -num_actions:], experience.values, experience.advantages, action_mask=experience.action_mask
+            values[:, -num_actions:],
+            experience.values,
+            experience.advantages,
+            action_mask=experience.action_mask if self.apply_loss_mask else None,
         )
         critic_loss = critic_loss * self.vf_coef
         self.critic_booster.backward(loss=critic_loss, optimizer=self.critic_optim)