Merge branch 'master' into dev/pagolnar/ex_imagenet

README.md

@@ -8,14 +8,19 @@ This folder contains end-to-end applications that use DeepSpeed to train and use
 There are several training and finetuning examples so please see the individual folders for specific instructions.
 
 ## 3. Inference
-The DeepSpeed Huggingface inference [README](./inference/huggingface/README.md) explains how to get started with running DeepSpeed Huggingface inference examples.
+- The DeepSpeed-MII inference [README](./inference/mii/README.md) explains how to get started with running model inference with [DeepSpeed-MII](https://github.com/Microsoft/DeepSpeed-MII) and [DeepSpeed-FastGen](https://github.com/microsoft/DeepSpeed/tree/master/blogs/deepspeed-fastgen).
+- The DeepSpeed Huggingface inference [README](./inference/huggingface/README.md) explains how to get started with running DeepSpeed Huggingface inference examples.
 
 ## 4. Compression
 Model compression examples.
 
 ## 5. Benchmarks
 All benchmarks that use the DeepSpeed library are maintained in this folder.
 
+# Build Pipeline Status
+| Description | Status |
+| ----------- | ------ |
+| Integrations | [![nv-ds-chat](https://github.com/microsoft/DeepSpeed/actions/workflows/nv-ds-chat.yml/badge.svg?branch=master)](https://github.com/microsoft/DeepSpeed/actions/workflows/nv-ds-chat.yml) |
 
 # Contributing
 

applications/DeepSpeed-Chat/.gitignore

@@ -0,0 +1,137 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+results/
+outputs/
+
+.amltconfig
+.test_output
+*.hdf5
+*.h5

applications/DeepSpeed-Chat/training/step3_rlhf_finetuning/ppo_trainer.py → applications/DeepSpeed-Chat/dschat/rlhf/ppo_trainer.py

@@ -4,20 +4,18 @@
 # DeepSpeed Team
 import torch
 import torch.nn.functional as F
-import sys
-import os
+import time
 import deepspeed
 from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+from deepspeed.accelerator import get_accelerator
 
-sys.path.append(
-    os.path.abspath(os.path.join(os.path.dirname(__file__), os.path.pardir)))
-
-from utils.utils import print_rank_0
+from dschat.utils.utils import print_rank_0
 
 
 def print_all_ranks(tag, value, rank):
     world_size = torch.distributed.get_world_size()
-    all_tensor = torch.zeros(world_size, dtype=torch.float32).cuda()
+    all_tensor = torch.zeros(world_size, dtype=torch.float32).to(
+        get_accelerator().current_device_name())
     all_tensor[rank] = value
     torch.distributed.all_reduce(all_tensor, op=torch.distributed.ReduceOp.SUM)
     print_rank_0(f'{tag} {all_tensor}', rank)
@@ -57,6 +55,12 @@ def __init__(self, rlhf_engine, args):
         self.end_of_conversation_token_id = self.tokenizer(
             args.end_of_conversation_token)['input_ids'][-1]
         self.z3_enabled = args.actor_zero_stage == 3
+        self.compute_fp32_loss = self.args.compute_fp32_loss
+
+        # In case the generated experience is not valid (too short), we use the last valid
+        # generated experience. Alternatively, we can skip the step (on all workers).
+        # For now, use the last valid experience which is a simpler solution
+        self.last_generated_experience = None
 
         # Those value can be changed
         self.kl_ctl = 0.1
@@ -65,18 +69,28 @@ def __init__(self, rlhf_engine, args):
         self.cliprange_value = 0.2
         self.gamma = 1.0
         self.lam = 0.95
+        self.generate_time = 0.0
 
     def _generate_sequence(self, prompts, mask, step):
 
         max_min_length = self.max_answer_seq_len + prompts.shape[1]
 
+        # This has been added due to a probability/nan error that happens after
+        # meta-llama/Llama-2-7b-hf enabled do_sample:
+        # https://huggingface.co/meta-llama/Llama-2-7b-hf/commit/6fdf2e60f86ff2481f2241aaee459f85b5b0bbb9
+        if self.actor_model.module.config.model_type == "llama":
+            kwargs = dict(do_sample=False)
+        else:
+            kwargs = dict()
+
         with torch.no_grad():
             seq = self.actor_model.module.generate(
                 prompts,
                 attention_mask=mask,
                 max_length=max_min_length,
                 pad_token_id=self.tokenizer.pad_token_id,
-                synced_gpus=self.z3_enabled)
+                synced_gpus=self.z3_enabled,
+                **kwargs)
 
         # Filter out seq with no answers (or very short). This happens when users directly use the pre-training ckpt without supervised finetuning
         # NOTE: this will causes each GPU has different number of examples
@@ -86,7 +100,8 @@ def _generate_sequence(self, prompts, mask, step):
         ans = seq[:, prompt_length:]
         valid_ans_len = (ans != self.tokenizer.pad_token_id).sum(dim=-1)
 
-        if self.args.print_answers:
+        if self.args.print_answers and (step % self.args.print_answers_interval
+                                        == 0):
             print(
                 f"--- prompt --> step={step}, rank={torch.distributed.get_rank()}, {self.tokenizer.batch_decode(prompts, skip_special_tokens=True)}"
             )
@@ -98,16 +113,38 @@ def _generate_sequence(self, prompts, mask, step):
         for i in range(batch_size):
             if valid_ans_len[
                     i] <= 1:  # if the answer is shorter than 1 token, drop it
+                print(
+                    f'Dropping too short generated answer: {step=}: \n'
+                    f'prompts: {self.tokenizer.batch_decode(prompts, skip_special_tokens=False)}\n'
+                    f'answers: {self.tokenizer.batch_decode(ans, skip_special_tokens=False)}'
+                )
                 continue
             else:
                 out_seq.append(seq[i:i + 1])
-        out_seq = torch.cat(out_seq, dim=0)  # concate output in the batch dim
+
+        if not out_seq:
+            print(
+                f'All generated results are too short for rank={self.args.local_rank} step={step}\n'
+                f'-> prompts: {self.tokenizer.batch_decode(prompts, skip_special_tokens=False)}\n'
+                f'-> answers: {self.tokenizer.batch_decode(ans, skip_special_tokens=False)}'
+            )
+            return None
+
+        out_seq = torch.cat(out_seq, dim=0)  # concat output in the batch dim
 
         return out_seq
 
     def generate_experience(self, prompts, mask, step):
         self.eval()
+        generate_start = time.time()
         seq = self._generate_sequence(prompts, mask, step)
+        generate_end = time.time()
+        if seq is None:
+            assert self.last_generated_experience is not None, f'Invalid generated experience at {step=}'
+            prompts = self.last_generated_experience['prompts']
+            seq = self.last_generated_experience['seq']
+        else:
+            self.last_generated_experience = {'prompts': prompts, 'seq': seq}
         self.train()
 
         pad_token_id = self.tokenizer.pad_token_id
@@ -124,6 +161,11 @@ def generate_experience(self, prompts, mask, step):
 
         logits = output.logits
         logits_ref = output_ref.logits
+        if self.compute_fp32_loss:
+            logits = logits.to(torch.float)
+            logits_ref = logits_ref.to(torch.float)
+
+        self.generate_time = generate_end - generate_start
 
         return {
             'prompts': prompts,
@@ -226,6 +268,17 @@ def train_rlhf(self, inputs):
 
         return actor_loss, critic_loss
 
+    def get_overflow(self):
+        # Overflow is not expected when using bf16
+        # Therefore, DeepSpeed's BF16_Optimizer does not maintain an overflow indication
+        if self.args.dtype == "bf16":
+            return False, False
+
+        actor_overflow = self.actor_model.optimizer.overflow
+        critic_overflow = self.critic_model.optimizer.overflow
+
+        return actor_overflow, critic_overflow
+
     def actor_loss_fn(self, logprobs, old_logprobs, advantages, mask):
         ## policy gradient loss
         log_ratio = (logprobs - old_logprobs) * mask
@@ -243,6 +296,9 @@ def critic_loss_fn(self, values, old_values, returns, mask):
             old_values - self.cliprange_value,
             old_values + self.cliprange_value,
         )
+        if self.compute_fp32_loss:
+            values = values.float()
+            values_clipped = values_clipped.float()
         vf_loss1 = (values - returns)**2
         vf_loss2 = (values_clipped - returns)**2
         vf_loss = 0.5 * torch.sum(