fix: multi-layer eagle forward hook bugs

new: unit test to check the qkv shape and forward_hook

Signed-off-by: Chenhan Yu <[email protected]>

Author: ChenhanYu

modelopt/torch/speculative/plugins/megatron_eagle.py

@@ -524,9 +524,11 @@ def __init__(
         if self._num_aux_hidden_states > 0:
             # Register forward hook to the last EAGLE3 layer to extract the pre-norm hidden_state
             # for eagle3 auto regression.
-            layer = self.decoder.layers[-1]
-            layer.register_forward_hook(self._eagle3_layer_forward_hook)
+            last_layer = self.decoder.layers[-1]
+            last_layer.register_forward_hook(self._eagle3_layer_forward_hook)
 
+            # The first EAGLE3 layer needs to be specialized.
+            layer = self.decoder.layers[0]
             self_attention = layer.self_attention
             if not isinstance(self_attention, SelfAttention):
                 raise ValueError("EAGLE-3 only support SelfAttention (MHA, GQA).")

tests/gpu/torch/speculative/plugins/test_speculative_megatron_modules.py

@@ -32,6 +32,12 @@
 from modelopt.torch.speculative.plugins.megatron_medusa import _DynamicMedusaGPTModel
 from modelopt.torch.speculative.utils import Tree, get_default_attention_mask_and_position_ids
 
+ALGO_TO_CONFIG = {
+    "eagle1": mtsp.config.EAGLE1_DEFAULT_CFG,
+    "eagle3": mtsp.config.EAGLE3_DEFAULT_CFG,
+    "eagle-mtp": mtsp.config.EAGLE_MTP_DEFAULT_CFG,
+}
+
 
 def _test_speculative_gpt_model(
     algo, num_medusa_heads_or_eagle_layers, activation_func, normalization, rank, size
@@ -64,18 +70,42 @@ def _test_speculative_gpt_model(
 
         # Type checking
         assert isinstance(model, _DynamicMedusaGPTModel)
-    elif algo == "eagle":
-        config = {"eagle_architecture_config": deepcopy(default_eagle_config)}
-        config["eagle_architecture_config"]["hidden_size"] = model.config.hidden_size
-        config["eagle_architecture_config"]["vocab_size"] = model.vocab_size
-        config["eagle_architecture_config"]["draft_vocab_size"] = model.vocab_size
+    elif algo in {"eagle1", "eagle3"}:
+        mtsp_config = ALGO_TO_CONFIG[algo]
+
+        mtsp_config["config"]["eagle_architecture_config"]["num_hidden_layers"] = (
+            num_medusa_heads_or_eagle_layers
+        )
+        mtsp_config["config"]["eagle_architecture_config"]["hidden_size"] = model.config.hidden_size
+        mtsp_config["config"]["eagle_architecture_config"]["vocab_size"] = model.vocab_size
+        mtsp_config["config"]["eagle_architecture_config"]["draft_vocab_size"] = model.vocab_size
 
-        model = mtsp.convert(model, [("eagle", config)])
+        model = mtsp.convert(model, mtsp_config)
 
         # Type checking
         assert isinstance(model, _DynamicEagleGPTModel)
     else:
-        raise ValueError("Only algo={eagle, medusa} are supported!")
+        raise ValueError("Only algo={eagle1, eagle3, medusa} are supported!")
+
+    if algo == "eagle3":
+        first_layer = model.eagle_module.decoder.layers[0]
+        last_layer = model.eagle_module.decoder.layers[-1]
+        # Eagle3 QKV input_dim is 2x of hidden_size
+        assert (
+            first_layer.self_attention.linear_qkv.weight.shape[-1] == model.config.hidden_size * 2
+        )
+        # Eagle3 attention has a forward_pre_hook to handle additional features to be concatenated
+        assert len(first_layer.self_attention._forward_pre_hooks) > 0
+        # Eagle3 last layer has a forward hook to extrat the pre_norm hidden_state
+        assert len(last_layer._forward_hooks) > 0
+    elif algo == "eagle1":
+        first_layer = model.eagle_module.decoder.layers[0]
+        last_layer = model.eagle_module.decoder.layers[-1]
+        # Eagle1 QKV input_dim the same as hidden_size
+        assert first_layer.self_attention.linear_qkv.weight.shape[-1] == model.config.hidden_size
+        # No forward_hook or forward_pre_hook are needed
+        assert len(first_layer.self_attention._forward_pre_hooks) == 0
+        assert len(last_layer._forward_hooks) == 0
 
     # Bfloat16
     model = model.to(torch.bfloat16)
@@ -104,7 +134,7 @@ def _test_speculative_gpt_model(
 
         assert medusa_loss.shape[0] == batch_size
         assert medusa_loss.shape[1] == max_sequence_length
-    elif algo == "eagle":
+    elif algo in {"eagle1", "eagle3"}:
         labels = torch.randint(0, vocab_size, (batch_size, max_sequence_length)).cuda()
         eagle_loss = model(prompt_tokens, position_ids, attention_mask, labels=labels)
 
@@ -115,8 +145,10 @@ def _test_speculative_gpt_model(
 @pytest.mark.parametrize(
     ("algo", "num_medusa_heads_or_eagle_layers", "activation_func", "normalization"),
     [
-        ("eagle", 1, "squared_relu", "LayerNorm"),  # MHA
-        ("eagle", 2, "swiglu", "RMSNorm"),  # GQA
+        ("eagle1", 1, "squared_relu", "LayerNorm"),  # MHA
+        ("eagle1", 2, "swiglu", "RMSNorm"),  # GQA
+        ("eagle3", 1, "swiglu", "RMSNorm"),  # GQA
+        ("eagle3", 2, "swiglu", "RMSNorm"),  # GQA
         ("medusa", 1, "squared_relu", "LayerNorm"),  # MHA
         ("medusa", 2, "swiglu", "RMSNorm"),  # GQA
     ],