Draft
Conversation
rob-maron
commented
Mar 18, 2026
Comment on lines
+47
to
+64
|
|
||
|
|
||
| from causal_conv1d import causal_conv1d_fn as _cuda_causal_conv1d_fn | ||
|
|
||
|
|
||
| def causal_conv1d_fn( | ||
| x: torch.Tensor, | ||
| weight: torch.Tensor, | ||
| bias: torch.Tensor | None = None, | ||
| activation: str | None = None, | ||
| seq_idx: torch.Tensor | None = None, | ||
| **kwargs, | ||
| ) -> tuple[torch.Tensor, None]: | ||
| # FLA convention is [B, T, D]; causal_conv1d expects [B, D, T] | ||
| out = _cuda_causal_conv1d_fn( | ||
| x.transpose(1, 2), weight, bias=bias, seq_idx=seq_idx, activation=activation | ||
| ) | ||
| return out.transpose(1, 2), None |
Author
There was a problem hiding this comment.
[when launching a multinode training run using submit_sft.py]:
With FLA's causal_conv1d layer I was seeing a few issues:
- CUDA illegal memory access in
fla/modules/conv/triton/ops.pyduring the backward pass torch.utils.checkpoint.CheckpointError: torch.utils.checkpoint: Recomputed values for the following tensors have different metadata than during the forward pass.
I replaced it the layer with one from causal_conv1d, which fixes both issues. Lmk if there is a preferred other way around it
| nodes_array=($nodes) | ||
| head_node=${nodes_array[0]} | ||
| export head_node_ip=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address) | ||
| export head_node_ip=$(getent hosts "$head_node" | awk '{print $1}') |
Author
There was a problem hiding this comment.
The srun here was deadlocking, this change makes it resolve locally
Comment on lines
224
to
+293
| @@ -239,11 +239,7 @@ def _to_hf_qwen3next(self, state_dict: dict[str, Any]) -> dict[str, Any]: | |||
| ) | |||
| hf_state_dict.update(local_expert_fqn) | |||
| else: | |||
| n_experts = ( | |||
| self.model_args.moe_args.num_experts | |||
| if "shared" not in key | |||
| else self.model_args.moe_args.num_shared_experts | |||
| ) | |||
| n_experts = self.model_args.moe_args.num_experts | |||
| split_values = self._split_experts_weights(value, n_experts) | |||
| for expert_num in range(n_experts): | |||
| new_key = new_abstract_key.format(layer_num, expert_num) | |||
| @@ -268,15 +264,9 @@ def _from_hf_qwen3next(self, hf_state_dict: dict[str, Any]) -> dict[str, Any]: | |||
| expert_weights_by_layer: dict[str, dict[str, dict]] = {} | |||
|
|
|||
| for key, value in hf_state_dict.items(): | |||
| if "mlp.experts." in key or "mlp.shared_expert." in key: | |||
| abstract_key = re.sub( | |||
| r"(\d+)", "{}", key, count=2 if "experts" in key else 1 | |||
| ) | |||
| if "experts" in key: | |||
| layer_num, expert_num = re.findall(r"\d+", key) | |||
| else: | |||
| layer_num = re.search(r"\d+", key).group(0) | |||
| expert_num = None | |||
| if "mlp.experts." in key and "mlp.shared_expert." not in key: | |||
| abstract_key = re.sub(r"(\d+)", "{}", key, count=2) | |||
| layer_num, expert_num = re.findall(r"\d+", key) | |||
| titan_abstract_key = self.from_hf_map[abstract_key] | |||
| new_key = titan_abstract_key.format(layer_num) | |||
|
|
|||
| @@ -285,7 +275,7 @@ def _from_hf_qwen3next(self, hf_state_dict: dict[str, Any]) -> dict[str, Any]: | |||
| if titan_abstract_key not in expert_weights_by_layer[layer_num]: | |||
| expert_weights_by_layer[layer_num][titan_abstract_key] = {} | |||
| expert_weights_by_layer[layer_num][titan_abstract_key][ | |||
| expert_num | |||
| int(expert_num) | |||
| ] = value | |||
|
|
|||
| if isinstance(value, DTensor): | |||
| @@ -296,16 +286,11 @@ def _from_hf_qwen3next(self, hf_state_dict: dict[str, Any]) -> dict[str, Any]: | |||
| value.device_mesh, | |||
| ) | |||
| else: | |||
| n_experts = ( | |||
| self.model_args.moe_args.num_experts | |||
| if "experts" in key | |||
| else self.model_args.moe_args.num_shared_experts | |||
| ) | |||
| stacked_value = self._concatenate_expert_weights( | |||
| expert_weights_by_layer, | |||
| titan_abstract_key, | |||
| layer_num, | |||
| n_experts, | |||
| self.model_args.moe_args.num_experts, | |||
Author
There was a problem hiding this comment.
The shared experts were in the grouped expert path (which expects 3d tensors), causing IndexError: too many indices for tensor of dimension 2. This takes them out
dmahan93
requested changes
Mar 18, 2026
There was a problem hiding this comment.
we should move these to a cluster specific setup as your settings conflict with the B200 cluster, maybe some host name based .sh routing?
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.