TP support for reverse KL loss

fast_llm/engine/training/config.py

@@ -361,7 +361,6 @@ def _validate(self) -> None:
             # TODO: Add support.
             Assert.eq(self.model.distributed.pipeline_parallel, 1)
             # TODO: Check if these work.
-            Assert.eq(self.model.distributed.tensor_parallel, 1)
             Assert.eq(self.model.distributed.sequence_data_parallel, 1)
         if self.run.experiment_dir is None:
             assert not self.training.checkpoint.enabled()

fast_llm/functional/cross_entropy.py

@@ -141,17 +141,23 @@ def _fused_cross_entropy_forward_backward(
         predicted_logits = logits_norm.gather(1, target)
         if group is not None:
             predicted_logits = target_mask * predicted_logits
+
             all_reduce(predicted_logits, op=ReduceOp.SUM, group=group)
     else:
         predicted_logits = (target * logits_norm).sum(dim=-1, keepdim=True)
+    if group is not None and target_format != TargetFormat.labels:
+        # this is needed because on each rank we calculate log Z - sum_i t_i * z_i, where z_i is logit.
+        # Then we average on line 160: 1/K sum_ranks (log Z - sum_i t_i * z_i)
+        # = log Z - 1/K sum_ranks (sum_i t_i * z_i), where is the global predicted_logits, so without multiplying it by K 1/K there does not cancel out.
+        predicted_logits = predicted_logits * group.size()
 
     per_sample_loss = sum_exp_logits.log() - predicted_logits
     if loss_mask is not None:
         per_sample_loss = per_sample_loss * loss_mask
 
     loss = per_sample_loss.mean()
     if target_format != TargetFormat.labels and group is not None:
-        all_reduce(loss, op=ReduceOp.MEAN, group=group)
+        all_reduce(loss, op=ReduceOp.AVG, group=group)
 
     return loss, grad
 
@@ -213,71 +219,72 @@ def cross_entropy_forward_backward(
         )
 
 
+def distributed_log_softmax(
+    logits: torch.Tensor, logits_scale_factor: float = 1.0, group: ProcessGroup | None = None, dim: int = -1
+):
+    logits_norm, _, sum_exp_logits = _fused_softmax_base(logits, logits_scale_factor, group=group, dim=dim)
+
+    return logits_norm - sum_exp_logits.log()  # log_softmax
+
+
 def _torch_reverse_kl_forward_backward(
     logits: torch.Tensor,
     target: torch.Tensor,
     loss_mask: torch.Tensor | None,
     grad_output: float | None,
-    logits_scale_factor: float,
     target_format: TargetFormat,
     group: ProcessGroup | None = None,
+    logits_scale_factor: float = 1.0,
     teacher_softmax_temperature: float = 1.0,
+    **kwargs,
 ) -> tuple[torch.Tensor, torch.Tensor | None]:
     """
     Reverse KL using PyTorch's native kl_div function.
-    Much simpler and more reliable than custom implementation!
+    This is used for TP version where we split accross vocab dimantion. KL is additive over partitions of the vocab.
+
+    Takes:
+        logits: [BxS, V] or [B, S, V]
+        target: [BxS, V] or [B, S, V] (logits format)
+        loss_mask: [BxS] or [B, S] or None
+        ...
     """
-    Assert.eq(target_format, TargetFormat.logits, msg="Reverse KL only supports logits format")
+    Assert.eq(
+        teacher_softmax_temperature,
+        1,
+        msg="Teacher softmax temperature must be 1 for sequence-tensor-parallel reverse KL",
+    )
+    Assert.eq(logits_scale_factor, 1, msg="Logits scale factor must be 1 for sequence-tensor-parallel reverse KL")
     Assert.eq(target.shape, logits.shape)
     assert target.dtype.is_floating_point, target.dtype
     if loss_mask is not None:
         Assert.eq(loss_mask.shape, logits.shape[:-1])
 
-    # Compute log probabilities - let _fused_softmax handle scaling internally
-    # teacher_probs = _fused_softmax(target, logits_scale_factor * (1 / teacher_softmax_temperature), group)
-    # # teacher_log_probs = torch.log(teacher_probs + 1e-8)  # log(p)
-    # teacher_probs = torch.clamp(teacher_probs, min=1e-7)  # or even 1e-6
-    # teacher_log_probs = torch.log(teacher_probs)
-
-    # Scale target logits more carefully
-    scaled_target = target * (logits_scale_factor / teacher_softmax_temperature)
-
-    # Clamp to prevent extreme values before log_softmax
-    scaled_target = torch.clamp(scaled_target, min=-50, max=50)
-    teacher_log_probs = torch.log_softmax(scaled_target, dim=-1)
-
-    # For reverse KL: KL(q||p) = Σ q * log(q/p) = Σ q * (log(q) - log(p))
-    # Use kl_div with: input=log(p), target=q, log_target=False
-    # This gives: Σ q * (log(q) - log(p)) = exactly what we want!
-
+    # Compute log probabilities
+    teacher_log_probs = distributed_log_softmax(target.float(), group=group)
+    # batch_size = logits.shape[0]
     with torch.enable_grad():
-        logits_ = logits.detach().requires_grad_(grad_output is not None)
-
-        # Use log_softmax for consistency instead of _fused_softmax
-        scaled_logits = logits_ * logits_scale_factor
-        scaled_logits = torch.clamp(scaled_logits, min=-50, max=50)
-        student_log_probs = torch.log_softmax(scaled_logits, dim=-1)
-
-        # Convert to probabilities for kl_div
-        # student_probs_ = torch.exp(student_log_probs)
+        logits_ = logits.float().detach().requires_grad_(grad_output is not None)
+        student_log_probs = distributed_log_softmax(logits_, group=group)
 
         # Reverse KL: input=teacher_log_probs, target=student_probs
-        if loss_mask is None:
-            loss = torch.nn.functional.kl_div(
-                teacher_log_probs,  # input = log(p)
-                student_log_probs,  # target = log(q)
-                reduction="batchmean",
-                log_target=True,
-            )
+        loss_terms = torch.nn.functional.kl_div(
+            teacher_log_probs,  # input = log(p)
+            student_log_probs,  # target = log(q)
+            reduction="none",
+            log_target=True,
+        ).sum(dim=-1)
+        if loss_mask is not None:
+            # loss mask is the same on all ranks for TP over vocab.
+            valid = loss_mask.to(loss_terms.dtype)
+            loss_terms = loss_terms * valid
+            valid_tokens = torch.tensor(valid.sum(), device=loss_terms.device, dtype=loss_terms.dtype)
         else:
-            # Apply loss mask - this requires some reshaping
-            loss_per_sample = torch.nn.functional.kl_div(
-                teacher_log_probs, student_log_probs, reduction="none", log_target=True
-            ).sum(dim=-1)
-            loss = (loss_per_sample * loss_mask).mean()
+            valid_tokens = torch.prod(torch.tensor(loss_terms.shape, device=loss_terms.device, dtype=loss_terms.dtype))
+        loss = loss_terms.sum()  # sums over batch and seq. len.
 
-        if group is not None and target_format != TargetFormat.labels:
-            all_reduce(loss, op=ReduceOp.MEAN, group=group)
+        if group is not None:
+            all_reduce(loss, op=ReduceOp.SUM, group=group)
+        loss /= valid_tokens
 
         if grad_output is not None:
             loss.backward(torch.full_like(loss, grad_output))
@@ -297,6 +304,7 @@ def reverse_kl_forward_backward(
     logits_scale_factor: float = 1.0,
     teacher_softmax_temperature: float = 1.0,
     target_format: TargetFormat = TargetFormat.labels,
+    sequence_parallel_logits: bool = False,
 ) -> tuple[torch.Tensor, torch.Tensor | None]:
     """
     Compute reverse KL divergence: KL(q||p) where q is the predicted distribution (student) and p is the target (teacher).
@@ -309,37 +317,36 @@ def reverse_kl_forward_backward(
     - Standard CE: KL(p||q) = mode-covering (spreads mass broadly)
     - Reverse KL: KL(q||p) = mode-seeking (focuses on target modes)
 
-    Args:
-        logits: Model predictions [batch_size, ..., vocab_size]
-        target: Target distribution or labels
-        loss_mask: Optional mask for loss computation
-        grad_output: Gradient output scale factor
-        group: Process group for tensor parallelism
-        logits_scale_factor: Temperature scaling factor (1/T)
-        target_format: Format of target (labels or logits)
+    Takes:
+        logits: [BxS, V] or [B, S, V], where V is local vocab size
+        target: [BxS, V] or [B, S, V] (logits format)
+        loss_mask: [BxS] or [B, S] or None
+        ...
 
     Returns:
         loss: Reverse KL divergence loss
         grad: Gradients w.r.t. logits
-
-    Example usage:
-        # Replace standard cross-entropy with reverse KL
-        # loss, grad = cross_entropy_forward_backward(logits, target, ...)
-        loss, grad = reverse_kl_forward_backward(logits, target,
-                                               loss_mask=None,
-                                               grad_output=1.0,
-                                               logits_scale_factor=1.0,
-                                               target_format=TargetFormat.labels)
     """
-    if target_format == TargetFormat.labels:
-        Assert.eq(target.shape, logits.shape[:-1])
-        Assert.eq(target.dtype, torch.int64)
-    else:
-        Assert.eq(target.shape, logits.shape)
-        assert target.dtype.is_floating_point, target.dtype
-        if loss_mask is not None:
-            Assert.eq(loss_mask.shape, logits.shape[:-1])
+
+    if sequence_parallel_logits:
+        # TODO: see hybrid dev branch where it is implemented
+        raise NotImplementedError("Sequence-parallel reverse KL is not implemented yet, set vocab_parallel true")
+
+    Assert.eq(target_format, TargetFormat.logits, msg="Reverse KL only supports logits format")
+    Assert.eq(target.shape, logits.shape)
+    assert target.dtype.is_floating_point, target.dtype
+    if loss_mask is not None:
+        Assert.eq(loss_mask.shape, logits.shape[:-1])
+
     # TODO: implement fused?
-    return _torch_reverse_kl_forward_backward(
-        logits, target, loss_mask, grad_output, logits_scale_factor, target_format, group, teacher_softmax_temperature
+    distillation_loss, distillation_grad = _torch_reverse_kl_forward_backward(
+        logits=logits,
+        target=target,
+        loss_mask=loss_mask,
+        grad_output=grad_output,
+        logits_scale_factor=logits_scale_factor,
+        target_format=target_format,
+        teacher_softmax_temperature=teacher_softmax_temperature,
+        group=group,
     )
+    return distillation_loss, distillation_grad

fast_llm/layers/language_model/head.py

@@ -387,7 +387,9 @@ def _logits_cross_entropy_forward_backward(
                     target_format=(
                         TargetFormat.labels if self._config.distillation_model is None else TargetFormat.logits
                     ),
+                    sequence_parallel_logits=self._sequence_parallel_logits,
                 )
+
             elif self._config.distillation_loss_implementation == DistillationLossImpl.cross_entropy:
                 distillation_loss, distillation_grad = cross_entropy_forward_backward(
                     logits.flatten(0, -2),

fast_llm/models/gpt/conversion/llama.py

@@ -198,19 +198,19 @@ def import_config(cls, config: dict) -> dict:
         elif rope_type == "llama3":
             rotary_config.update(
                 {
-                    "scale_factor": config["factor"],
-                    "low_frequency_factor": config["low_freq_factor"],
-                    "high_frequency_factor": config["high_freq_factor"],
-                    "original_context_length": config["original_max_position_embeddings"],
+                    "scale_factor": config["rope_scaling"]["factor"],
+                    "low_frequency_factor": config["rope_scaling"]["low_freq_factor"],
+                    "high_frequency_factor": config["rope_scaling"]["high_freq_factor"],
+                    "original_context_length": config["rope_scaling"]["original_max_position_embeddings"],
                 }
             )
         elif rope_type == "yarn":
             rotary_config.update(
                 {
-                    "attention_factor": config["attention_factor"],
-                    "beta_fast": config["beta_fast"],
-                    "beta_slow": config["beta_slow"],
-                    "original_context_length": config["original_max_position_embeddings"],
+                    "attention_factor": config["rope_scaling"]["attention_factor"],
+                    "beta_fast": config["rope_scaling"]["beta_fast"],
+                    "beta_slow": config["rope_scaling"]["beta_slow"],
+                    "original_context_length": config["rope_scaling"]["original_max_position_embeddings"],
                 }
             )
         else: