[TRTLLM-10060][feat] Enable attention dp for Nemotron Super v3.

Signed-off-by: nv-guomingz <[email protected]>

Author: nv-guomingz

tensorrt_llm/_torch/models/checkpoints/hf/nemotron_h_weight_mapper.py

@@ -11,7 +11,7 @@ class NemotronHHfWeightMapper(HfWeightMapper):
 
     def preprocess_weights(self, weights: dict) -> dict:
         config = self.config.pretrained_config
-        tp_size = self.config.mapping.tp_size
+        tp_size = 1 if self.config.mapping.enable_attention_dp else self.config.mapping.tp_size
         tp_rank = self.config.mapping.tp_rank
         d_inner = config.mamba_head_dim * config.mamba_num_heads
 

tensorrt_llm/_torch/models/modeling_nemotron_h.py

@@ -32,7 +32,7 @@
 from ..modules.decoder_layer import DecoderLayer
 from ..modules.embedding import Embedding
 from ..modules.fused_moe import MoEWeightLoadingMode, create_moe
-from ..modules.linear import Linear
+from ..modules.linear import Linear, TensorParallelMode
 from ..modules.mamba.mamba2_mixer import Mamba2Mixer
 from ..modules.mlp import MLP
 from ..modules.multi_stream_utils import maybe_execute_in_parallel
@@ -85,8 +85,10 @@ def __init__(
         self,
         model_config: ModelConfig[NemotronHConfig],
         layer_idx: int,
+        reduce_output: bool = False,
     ):
         config = model_config.pretrained_config
+
         super().__init__(
             hidden_size=config.hidden_size,
             num_attention_heads=config.num_attention_heads,
@@ -97,6 +99,7 @@ def __init__(
             layer_idx=layer_idx,
             dtype=config.torch_dtype,
             config=model_config,
+            reduce_output=reduce_output,
         )
 
     def forward(
@@ -154,6 +157,7 @@ def __init__(
             shared_expert_intermediate_size = (
                 config.moe_shared_expert_intermediate_size *
                 config.n_shared_experts)
+
             self.shared_experts = MLP(
                 hidden_size=config.hidden_size,
                 intermediate_size=shared_expert_intermediate_size,
@@ -163,7 +167,8 @@ def __init__(
                 config=model_config,
                 layer_idx=self.layer_idx,
                 reduce_output=False,
-            )
+                overridden_tp_size=1
+                if model_config.mapping.enable_attention_dp else None)
         # Setup MoE gate.
         self.gate = DeepseekV3Gate(
             self.hidden_size,
@@ -193,11 +198,14 @@ def __init__(
             activation_type=self.activation_type,
         )
 
-        # AllReduce for combining shared and routed expert outputs in multi-GPU settings.
-        self.allreduce = AllReduce(
-            mapping=model_config.mapping,
-            strategy=model_config.allreduce_strategy,
-        )
+        if not model_config.mapping.enable_attention_dp:
+            # AllReduce for combining shared and routed expert outputs in multi-GPU settings.
+            self.allreduce = AllReduce(
+                mapping=model_config.mapping,
+                strategy=model_config.allreduce_strategy,
+            )
+        else:
+            self.allreduce = None
 
         # Setup latent projection layers.
         # These layers should NOT be TP-sharded to ensure MoE receives
@@ -322,7 +330,11 @@ def __init__(
         elif layer_type == "-":
             self.mixer = MLPLayer(model_config, layer_idx)
         elif layer_type == "*":
-            self.mixer = TransformerLayer(model_config, layer_idx)
+            self.mixer = TransformerLayer(
+                model_config,
+                layer_idx,
+                reduce_output=not model_config.mapping.enable_attention_dp
+                and model_config.mapping.tp_size > 1)
         elif layer_type == "E":
             self.mixer = NemotronHMOE(model_config,
                                       layer_idx=layer_idx,
@@ -365,12 +377,24 @@ def __init__(self, model_config: ModelConfig[NemotronHConfig]):
             aux_stream_list[2],
         }
 
-        # calculate embeddings
-        self.embed_tokens = Embedding(
-            config.vocab_size,
-            config.hidden_size,
-            dtype=config.torch_dtype,
-        )
+        if model_config.mapping.enable_attention_dp:
+            # When attention_dp is enabled, we cannot do all_reduce since
+            # the problem size of different ranks are different.
+            # So, we don't do parallelism here.
+            self.embed_tokens = Embedding(
+                config.vocab_size,
+                config.hidden_size,
+                dtype=config.torch_dtype,
+            )
+        else:
+            self.embed_tokens = Embedding(
+                config.vocab_size,
+                config.hidden_size,
+                dtype=config.torch_dtype,
+                mapping=model_config.mapping,
+                tensor_parallel_mode=TensorParallelMode.COLUMN,
+                gather_output=True,
+            )
 
         # create layers
         layers = []

tensorrt_llm/_torch/modules/mamba/mamba2_mixer.py

@@ -20,6 +20,7 @@
 from torch import nn
 
 from tensorrt_llm._torch.modules.mamba.mamba2_metadata import Mamba2Metadata
+from tensorrt_llm.mapping import Mapping
 
 from ...attention_backend import AttentionMetadata
 from ...model_config import ModelConfig
@@ -57,8 +58,20 @@ def __init__(
 
         config = config or ModelConfig()
         self.mapping = config.mapping
-        tp_rank = config.mapping.tp_rank
-        tp_size = config.mapping.tp_size
+
+        if config.mapping.enable_attention_dp:
+            self.mapping = Mapping(
+                world_size=config.mapping.pp_size,
+                tp_size=1,
+                pp_size=config.mapping.pp_size,
+                rank=config.mapping.rank,
+                gpus_per_node=config.mapping.gpus_per_node,
+                enable_attention_dp=True,
+            )
+            tp_size = 1
+        else:
+            self.mapping = config.mapping
+            tp_size = config.mapping.tp_size
 
         d_inner = head_dim * nheads
         d_in_proj = 2 * d_inner + 2 * n_groups * d_state + nheads
@@ -80,10 +93,6 @@ def __init__(
         self.remove_padding = remove_padding
         self.apply_silu = apply_silu
 
-        # tp
-        self.tp_size = tp_size
-        self.tp_rank = tp_rank
-
         # paged state parameters
         self.slot_mapping = None
         self.is_paged_state = False

tensorrt_llm/_torch/pyexecutor/mamba_cache_manager.py

@@ -45,7 +45,7 @@ def __init__(
         self.mamba_ssm_cache_dtype = ssm_cache_dtype
 
         # get tp size
-        tp_size = mapping.tp_size
+        tp_size = mapping.tp_size if not mapping.enable_attention_dp else 1
 
         # derive mamba parameters for conv and ssm states
         d_inner = head_dim * num_heads

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -5276,6 +5276,7 @@ class TestNemotronV3Super(LlmapiAccuracyTestHarness):
             (4, 4, False, True, True),
             (4, 1, False, False, True),
             (4, 4, True, False, True),
+            (4, 4, True, True, True),
             (4, 1, True, True, True),
             (4, 4, False, True, False),
             (4, 1, False, False, False),
@@ -5285,9 +5286,6 @@ class TestNemotronV3Super(LlmapiAccuracyTestHarness):
     )
     def test_auto_dtype_4gpus(self, tp_size, ep_size, attention_dp,
                               overlap_scheduler, cuda_graph):
-        if attention_dp:
-            pytest.skip(
-                "Attention DP is not supported for Nemotron-3-Super yet")
 
         kv_cache_config = KvCacheConfig(enable_block_reuse=False,
                                         mamba_ssm_cache_dtype="float32")
@@ -5314,7 +5312,18 @@ def test_auto_dtype_4gpus(self, tp_size, ep_size, attention_dp,
 
     @skip_pre_blackwell
     @pytest.mark.skip_less_mpi_world_size(8)
-    def test_nvfp4_8gpus(self):
+    @pytest.mark.parametrize(
+        "attention_dp",
+        [
+            False,
+            True,
+        ],
+        ids=[
+            "attention_dp_off",
+            "attention_dp_on",
+        ],
+    )
+    def test_nvfp4_8gpus(self, attention_dp):
         # Use this test to track the best performance config.
         # The optimized config is still under investigation.
         # Adding this test as placeholder.
@@ -5329,7 +5338,7 @@ def test_nvfp4_8gpus(self):
                 tensor_parallel_size=8,
                 moe_expert_parallel_size=8,
                 pipeline_parallel_size=1,
-                enable_attention_dp=False,
+                enable_attention_dp=attention_dp,
                 cuda_graph_config=CudaGraphConfig(max_batch_size=32,
                                                   enable_padding=True),
                 disable_overlap_scheduler=False,

tests/integration/test_lists/qa/llm_function_core.txt

@@ -247,7 +247,10 @@ accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-4
 accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-1-True-True-False]
 accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-1-False-False-True]
 accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-4-True-False-True]
-accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_nvfp4_8gpus
+accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-4-True-True-True]
+accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_nvfp4_8gpus[attention_dp_on]
+accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_nvfp4_8gpus[attention_dp_off]
+
 
 # multimodal accuracy tests
 accuracy/test_llm_api_pytorch_multimodal.py::TestQwen2_VL_7B::test_auto_dtype

tests/integration/test_lists/test-db/l0_dgx_b200.yml

@@ -23,7 +23,7 @@ l0_dgx_b200:
   - unittest/_torch/modules/test_fused_moe.py::test_fused_moe_mxfp4_mxfp8[enable_configurable_moe-True-8-64-TRTLLM]
   - unittest/_torch/modules/test_fused_moe.py::test_fused_moe_wfp4a16[enable_configurable_moe-TRTLLM-2880-dtype0]
   - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-4-False-True-True]
-  - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-1-True-True-True]
+  - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-4-True-True-True]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_trtllm-torch_compile=True]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_trtllm-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=False]
@@ -85,7 +85,7 @@ l0_dgx_b200:
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV32::test_nvfp4_multi_gpus[baseline] TIMEOUT (60)
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV32::test_nvfp4_multi_gpus[baseline_mtp1] TIMEOUT (60)
   - accuracy/test_disaggregated_serving.py::TestDeepSeekV32Exp::test_auto_dtype[False] TIMEOUT (60)
-  - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_nvfp4_8gpus TIMEOUT (60)
+  - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_nvfp4_8gpus[attention_dp_on] TIMEOUT (60)
 - condition:
     ranges:
       system_gpu_count:
@@ -163,6 +163,7 @@ l0_dgx_b200:
   - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-1-True-True-False]
   - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-1-False-False-True]
   - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-4-True-False-True]
+  - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-1-True-True-True]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_trtllm-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_cutlass-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_cutlass-torch_compile=True]