From 856545f8bbc0cf1dee38edf1c47b6cd3cd3b520f Mon Sep 17 00:00:00 2001
From: mrq <mrq@ecker.tech>
Date: Sat, 11 May 2024 22:23:29 -0500
Subject: [PATCH] nan loss detection (should have added it earlier), loss
 scaling for local backend + fp16

---
 vall_e/config.py            |  7 +++++++
 vall_e/engines/base.py      | 17 ++++++++++++++++-
 vall_e/engines/deepspeed.py |  7 +++++++
 3 files changed, 30 insertions(+), 1 deletion(-)

diff --git a/vall_e/config.py b/vall_e/config.py
index 128cbf1..f9ced76 100755
--- a/vall_e/config.py
+++ b/vall_e/config.py
@@ -544,6 +544,13 @@ class Trainer:
 			return torch.float8_e4m3fn
 		return torch.float32
 
+	@cached_property
+	def scale_loss(self):
+		# currently cannot feasibly apply loss scaling with DeepSpeed backend (it can handle it itself anyways)
+		if self.backend != "local":
+			return False
+		return self.dtype == torch.float16
+
 
 @dataclass()
 class Inference:
diff --git a/vall_e/engines/base.py b/vall_e/engines/base.py
index de598e0..66bc8a3 100755
--- a/vall_e/engines/base.py
+++ b/vall_e/engines/base.py
@@ -67,6 +67,9 @@ class Engine():
 
 		self._frozen_params = set()
 
+		self.max_nan_losses = 8
+		self.loss_scaler = torch.cuda.amp.GradScaler() if cfg.trainer.scale_loss else None
+
 	def freeze(self, freeze_all=True):
 		# set to freeze 
 		if self._cfg is None or not hasattr(self._cfg, "frozen_params"):
@@ -186,8 +189,11 @@ class Engine():
 		return self.module.forward(*args, **kwargs)
 
 	def backward(self, loss):
+		if self.loss_scaler is not None:
+			return self.loss_scaler.scale(loss / self.gradient_accumulation_steps).backward()
 		return (loss / self.gradient_accumulation_steps).backward()
 
+
 	def step(self):
 		with torch.set_grad_enabled(self.gradient_accumulation_steps > 1):
 			self.micro_steps += 1 
@@ -197,7 +203,11 @@ class Engine():
 				torch.nn.utils.clip_grad_norm_(self.module.parameters(), self.gradient_clipping)
 
 				self.global_steps += 1 
-				self.optimizer.step()
+				if self.loss_scaler is not None:
+					self.loss_scaler.step(self.optimizer)
+					self.loss_scaler.update()
+				else:
+					self.optimizer.step()
 				self.optimizer.zero_grad()
 
 				self._get_grad_norm()
@@ -232,6 +242,11 @@ class Engine():
 		losses = self.gather_attribute("loss")
 		loss = torch.stack([*losses.values()]).sum()
 
+		if torch.isnan(loss).any():
+			self.max_nan_losses = self.max_nan_losses - 1
+			if self.max_nan_losses < 0:
+				raise RuntimeError("Too many NaN losses detected.")
+
 		stats = {}
 		stats |= {k: v.item() for k, v in losses.items()}
 		stats |= self.gather_attribute("scalar")
diff --git a/vall_e/engines/deepspeed.py b/vall_e/engines/deepspeed.py
index 585eced..1ef8333 100755
--- a/vall_e/engines/deepspeed.py
+++ b/vall_e/engines/deepspeed.py
@@ -60,6 +60,8 @@ class Engine(DeepSpeedEngine):
 		self.global_samples = stats["global_samples"]
 		self.tokens_processed = stats["tokens_processed"]
 
+		self.max_nan_losses = 8
+
 	def freeze(self, freeze_all=True):
 		if self._cfg is None or not hasattr(self._cfg, "frozen_params"):
 			raise Exception("freeze_all=False yet self._cfg.frozen_params is None")
@@ -113,6 +115,11 @@ class Engine(DeepSpeedEngine):
 		losses = self.gather_attribute("loss")
 		loss = torch.stack([*losses.values()]).sum()
 
+		if torch.isnan(loss).any():
+			self.max_nan_losses = self.max_nan_losses - 1
+			if self.max_nan_losses < 0:
+				raise RuntimeError("Too many NaN losses detected.")
+
 		stats = {}
 		stats |= {k: v.item() for k, v in losses.items()}
 		stats |= self.gather_attribute("scalar")