Mixtral recipes

Author: trvachov

bionemo-recipes/models/mixtral/modeling_mixtral_te.py

@@ -279,9 +279,22 @@ def _restack_from_views(self) -> None:
         device = torch.cuda.current_device()
         for attr_name in ("experts_gate_up_weight", "experts_down_weight"):
             old_param = getattr(self, attr_name)
-            new_data = torch.empty_like(old_param, device=device)
-            torch.nn.init.normal_(new_data, mean=0.0, std=self.initializer_range)
-            setattr(self, attr_name, nn.Parameter(new_data))
+            if isinstance(old_param.data, DTensor):
+                # FSDP2 has sharded this param; materialize the local shard on CUDA
+                # and reconstruct the DTensor wrapper so FSDP2 can manage it.
+                local_data = old_param.data.to_local()
+                new_local = torch.empty(local_data.shape, dtype=local_data.dtype, device=device)
+                torch.nn.init.normal_(new_local, mean=0.0, std=self.initializer_range)
+                new_dtensor = DTensor.from_local(
+                    new_local,
+                    device_mesh=old_param.data.device_mesh,
+                    placements=old_param.data.placements,
+                )
+                setattr(self, attr_name, nn.Parameter(new_dtensor))
+            else:
+                new_data = torch.empty_like(old_param, device=device)
+                torch.nn.init.normal_(new_data, mean=0.0, std=self.initializer_range)
+                setattr(self, attr_name, nn.Parameter(new_data))
 
         # Re-sync views to point to the new stacked parameter
         self._sync_expert_views()
@@ -298,13 +311,15 @@ def _sync_expert_views(self) -> None:
         gate_up_w = self.experts_gate_up_weight
         if isinstance(gate_up_w, DTensor):
             gate_up_w = gate_up_w.to_local()
-        for i in range(self.num_local_experts):
+        num_local = gate_up_w.shape[0]
+        for i in range(num_local):
             object.__setattr__(self.experts_gate_up, f"weight{i}", gate_up_w[i])
 
         down_w = self.experts_down_weight
         if isinstance(down_w, DTensor):
             down_w = down_w.to_local()
-        for i in range(self.num_local_experts):
+        num_local_down = down_w.shape[0]
+        for i in range(num_local_down):
             object.__setattr__(self.experts_down, f"weight{i}", down_w[i])
 
     def set_ep_group(self, ep_group: dist.ProcessGroup, ep_mesh: DeviceMesh) -> None:
@@ -503,12 +518,20 @@ def __init__(
         self._fp8_recipe: transformer_engine.common.recipe.Recipe | None = fp8_recipe
         self._fp4_recipe: transformer_engine.common.recipe.Recipe | None = fp4_recipe
 
-        if fp8_recipe is not None and self.config.layer_precision is None:
-            if fp4_recipe is not None:
+        if self.config.layer_precision is None:
+            if fp8_recipe is not None and fp4_recipe is not None:
                 raise RuntimeError("Both FP8 and FP4 recipes provided, but no layer precision provided.")
-
-            warnings.warn("No layer precision provided, using FP8 recipe for all layers.", UserWarning)
-            self.config.layer_precision = ["fp8"] * self.config.num_hidden_layers
+            if fp8_recipe is not None:
+                warnings.warn("No layer precision provided, using FP8 recipe for all layers.", UserWarning)
+                self.config.layer_precision = ["fp8"] * self.config.num_hidden_layers
+            elif fp4_recipe is not None:
+                raise RuntimeError(
+                    "FP4 recipe provided but no layer_precision configured. "
+                    "Set layer_precision explicitly when using FP4."
+                )
+
+        if self.config.layer_precision is not None and "fp4" in self.config.layer_precision and fp4_recipe is None:
+            raise RuntimeError("layer_precision contains 'fp4' entries but no fp4_recipe was provided.")
 
         self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx, dtype=config.dtype)
 
@@ -857,6 +880,10 @@ def _unpad_input(hidden_states, attention_mask, unused_mask=None):
 class HFInferenceParams(InferenceParams):
     """Extension of the InferenceParams class to support HF generate() and beam search."""
 
+    # Required by transformers >= 5.4 _valid_auto_compile_criteria(); this
+    # custom TE-based cache is not compatible with torch.compile generate().
+    is_compileable = False
+
     def get_seq_length(self, layer_idx: int = 0) -> int:
         """Return the current cached sequence length.
 

bionemo-recipes/recipes/mixtral_native_te/Dockerfile

@@ -0,0 +1,9 @@
+# syntax=docker/dockerfile:1.4
+FROM nvcr.io/nvidia/pytorch:26.02-py3
+
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,source=requirements.txt,target=/requirements.txt \
+    PIP_CONSTRAINT= pip install -r /requirements.txt
+
+WORKDIR /workspace/bionemo
+COPY . .

bionemo-recipes/recipes/mixtral_native_te/README.md

@@ -0,0 +1,31 @@
+# TransformerEngine-accelerated Mixtral training with a native PyTorch training loop
+
+This folder demonstrates how to train TE-accelerated Mixtral with a native PyTorch training loop using FSDP2 for
+distributed training. The recipe mirrors the structure and conventions of `llama3_native_te`, and includes a Lingua-style
+configuration for natural-language pre-training on DCLM Baseline 1.0.
+
+## Commands
+
+Single GPU sanity run:
+
+```bash
+python train_fsdp2.py --config-name L0_sanity
+```
+
+Single GPU Lingua smoke run:
+
+```bash
+python train_fsdp2.py --config-name L2_lingua_8x1B num_train_steps=20 checkpoint.ckpt_dir=./checkpoints
+```
+
+Cluster or multi-GPU run:
+
+```bash
+torchrun --standalone --nproc_per_node=2 train_fsdp2.py --config-name L2_lingua_8x1B
+```
+
+## Notes
+
+- The Lingua config uses the `meta-llama/Meta-Llama-3-8B` tokenizer and streams `mlfoundations/dclm-baseline-1.0`.
+- `expert_parallel_size` remains `1` in this v1 recipe so it matches the existing Llama3 Lingua recipe structure.
+- Use `HF_TOKEN` for Hugging Face access and `WANDB_KEY` for Weights & Biases logging.