implement training for HPU

HPU cards (Gaudi 2 and 3) can't use Accelerate code path. This
contribution adds the training setup and loop for FSDP-only training.

Minor modifications required for HPUs specifically.

Signed-off-by: James Kunstle <[email protected]>

src/instructlab/training/constants.py

@@ -0,0 +1,9 @@
+SUPPORTED_MODEL_ARCHITECTURES: list[str] = [
+    "MistralForCausalLM",
+    "GPTDolomiteForCausalLM",
+    "LlamaForCausalLM",
+    "Starcoder2ForCausalLM",
+    "GemmaForCausalLM",
+    "MixtralForCausalLM",
+    "GraniteForCausalLM",
+]

src/instructlab/training/main_ds.py

@@ -35,6 +35,7 @@
     TorchrunArgs,
     TrainingArgs,
 )
+from instructlab.training.constants import SUPPORTED_MODEL_ARCHITECTURES
 from instructlab.training.multipack_sampler import (
     find_packing_max_batch_len_and_grad_accum,
 )
@@ -170,15 +171,9 @@ def setup_model(args, tokenizer, train_loader, grad_accum, flash_enabled):
         )
         model.config.eos_token_id = tokenizer.eos_token_id
 
-    assert model.__class__.__name__ in [
-        "MistralForCausalLM",
-        "GPTDolomiteForCausalLM",
-        "LlamaForCausalLM",
-        "Starcoder2ForCausalLM",
-        "GemmaForCausalLM",
-        "MixtralForCausalLM",
-        "GraniteForCausalLM",
-    ], f"Model class name: {model.__class__.__name__} is not supported."
+    assert (
+        model.__class__.__name__ in SUPPORTED_MODEL_ARCHITECTURES
+    ), f"Model class name: {model.__class__.__name__} is not supported."
 
     model = convert_loss_to_reduce_sum(model, use_dolomite=args.use_dolomite)
     model = add_noisy_embeddings(model, noise_alpha=args.NEFTune_alpha)
@@ -699,6 +694,8 @@ def check_hpu_compatible(
     """
     Using flash-attention (and by consequence, Dolomite models) is not supported
     if trying to train with Gaudi 2/3 cards.
+
+    Raises: RuntimeError
     """
 
     if using_hpu and any([using_dolomite, using_flash_attention]):
@@ -707,6 +704,27 @@ def check_hpu_compatible(
         )
 
 
+def train_hpu(
+    args,
+    model,
+    tokenizer,
+    data_loader,
+    grad_accum_steps,
+    metric_logger,
+):
+    # First Party
+    from instructlab.training.train_hpu_fsdp import main as main_hpu
+
+    main_hpu(
+        args=args,
+        model_name_or_path=model,
+        tokenizer=tokenizer,
+        grad_accum_steps=grad_accum_steps,
+        metric_logger=metric_logger,
+        data_loader=data_loader,
+    )
+
+
 def main(args):
     """
     Distributed training setup and execution.
@@ -815,8 +833,13 @@ def main(args):
         )
 
     else:
-        raise NotImplementedError(
-            "Training on Intel Gaudi 2/3 cards is not supported yet."
+        train_hpu(
+            args=args,
+            model=model,
+            tokenizer=tokenizer,
+            data_loader=train_loader,
+            grad_accum_steps=grad_accum,
+            metric_logger=metric_logger,
         )
 
     torch.distributed.barrier()