[TRTLLM-10171][fix] Correct attention handling in ModelConfig and KVCacheManager (#10330)

Signed-off-by: Jaedeok Kim <jaedeokk@nvidia.com>

tensorrt_llm/_torch/model_config.py

@@ -495,10 +495,15 @@ class ModelConfig(Generic[TConfig]):
         # TODO smor- currently assuming no rnn layers, no MOE
         from tensorrt_llm.bindings import ModelConfig as ModelConfigCpp
 
-        num_heads = self.pretrained_config.num_attention_heads // (
+        # Attention DP should not shard attention heads; use attn_tp_size=1 in that case
-            self.mapping.tp_size * self.mapping.cp_size)
+        # so downstream KV calculations see the full (non-partitioned) head count.
+        attn_tp_size = self.mapping.attn_tp_size if not self.mapping.enable_attention_dp else 1
+        attn_cp_size = self.mapping.attn_cp_size
 
-        hidden_size = self.pretrained_config.hidden_size // self.mapping.tp_size
+        num_heads = self.pretrained_config.num_attention_heads // (
+            attn_tp_size * attn_cp_size)
+
+        hidden_size = self.pretrained_config.hidden_size // attn_tp_size
 
         model_config_cpp = ModelConfigCpp(
             vocab_size=self.pretrained_config.vocab_size,
@@ -523,13 +528,12 @@ class ModelConfig(Generic[TConfig]):
         if isinstance(num_key_value_heads, (list, tuple)):
             # Per-layer KV heads (e.g., Nemotron-NAS, variable GQA models)
             num_kv_heads_per_layer = [
-                kv_heads // (self.mapping.tp_size * self.mapping.cp_size)
+                kv_heads // (attn_tp_size * attn_cp_size)
                 for kv_heads in num_key_value_heads
             ]
             model_config_cpp.num_kv_heads_per_layer = num_kv_heads_per_layer
         else:
-            num_kv_heads = num_key_value_heads // (self.mapping.tp_size *
+            num_kv_heads = num_key_value_heads // (attn_tp_size * attn_cp_size)
-                                                   self.mapping.cp_size)
             model_config_cpp.set_num_kv_heads(num_kv_heads)
 
         mlp_hidden_size = None

tensorrt_llm/_torch/pyexecutor/resource_manager.py

@@ -1138,7 +1138,8 @@ class KVCacheManager(BaseResourceManager):
             tensor_parallelism=self.mapping.tp_size,
             pipeline_parallelism=self.mapping.pp_size,
             rank=self.mapping.rank,
-            gpus_per_node=self.mapping.gpus_per_node)
+            gpus_per_node=self.mapping.gpus_per_node,
+            enable_attention_dp=self.mapping.enable_attention_dp)
 
         window_size_to_layers = self._get_window_size_to_layers()
         logger.debug(f"window_size_to_layers: {window_size_to_layers}")

tests/integration/test_lists/test-db/l0_a10.yml

@@ -17,6 +17,7 @@ l0_a10:
   - unittest/_torch/sampler/test_torch_sampler.py
   - unittest/_torch/sampler/test_torch_multi_arange.py
   - unittest/utils/test_util.py
+  - unittest/_torch/test_model_config.py
   - unittest/_torch/modeling/test_modeling_mistral.py
   - unittest/_torch/modeling/test_modeling_pixtral.py
   - unittest/_torch/sampler/test_trtllm_sampler.py

tests/unittest/_torch/test_model_config.py

@@ -0,0 +1,90 @@
+import types
+
+import pytest
+import torch
+
+from tensorrt_llm._torch.model_config import ModelConfig
+from tensorrt_llm.mapping import Mapping
+
+
+def make_pretrained_config(
+    *,
+    num_attention_heads: int = 16,
+    num_key_value_heads=8,
+    head_dim: int | None = None,
+    num_hidden_layers: int = 1,
+    vocab_size: int = 3000,
+):
+    # A minimal config object that provides the attributes used by
+    # ModelConfig.get_bindings_model_config().
+    hidden_size = head_dim * num_attention_heads
+    intermediate_size = hidden_size * 4
+
+    return types.SimpleNamespace(
+        architectures=["DummyArchitecture"],
+        num_attention_heads=num_attention_heads,
+        hidden_size=hidden_size,
+        intermediate_size=intermediate_size,
+        num_key_value_heads=num_key_value_heads,
+        head_dim=head_dim,
+        num_hidden_layers=num_hidden_layers,
+        vocab_size=vocab_size,
+        torch_dtype=torch.float16,
+    )
+
+
+@pytest.mark.parametrize(
+    "num_key_value_heads",
+    [
+        pytest.param(8, id="kv_heads_scalar"),
+        pytest.param([8, 20], id="kv_heads_per_layer_varied"),
+    ],
+)
+@pytest.mark.parametrize("enable_attention_dp", [False, True])
+@pytest.mark.parametrize(
+    "mapping_kwargs",
+    [
+        # Same tp/cp sizes, but different ways of setting attention TP:
+        # - No explicit attn_tp_size: Mapping infers it.
+        # - Explicit attn_tp_size: Mapping uses the provided value.
+        dict(world_size=8, tp_size=4, cp_size=2),
+        dict(world_size=4, tp_size=2, cp_size=2, attn_tp_size=4),
+    ],
+)
+def test_get_bindings_model_config_attention_dp_attn_tp_override(
+    enable_attention_dp, mapping_kwargs, num_key_value_heads
+):
+    mapping = Mapping(enable_attention_dp=enable_attention_dp, **mapping_kwargs)
+    cfg = make_pretrained_config(
+        # Keep values consistent:
+        # hidden_size = num_attention_heads * head_dim.
+        num_attention_heads=16,
+        head_dim=4,
+        num_key_value_heads=num_key_value_heads,
+        num_hidden_layers=2,
+    )
+    model_config = ModelConfig(pretrained_config=cfg, mapping=mapping)
+
+    tokens_per_block = 32
+    bindings_cfg = model_config.get_bindings_model_config(tokens_per_block=tokens_per_block)
+
+    # bindings hidden_size is sharded by attn_tp_size and attn_cp_size.
+    attn_tp_size = mapping.attn_tp_size if not mapping.enable_attention_dp else 1
+    attn_cp_size = mapping.attn_cp_size
+    assert bindings_cfg.num_heads == cfg.num_attention_heads // (attn_tp_size * attn_cp_size)
+    # bindings hidden_size is sharded by attn_tp_size.
+    assert bindings_cfg.hidden_size == cfg.hidden_size // attn_tp_size
+    if isinstance(cfg.num_key_value_heads, (list, tuple)):
+        expected_num_kv_heads_per_layer = [
+            kv // (attn_tp_size * attn_cp_size) for kv in cfg.num_key_value_heads
+        ]
+        assert list(bindings_cfg.num_kv_heads_per_layer) == expected_num_kv_heads_per_layer
+        assert bindings_cfg.num_kv_heads(0) == expected_num_kv_heads_per_layer[0]
+    else:
+        assert bindings_cfg.num_kv_heads(0) == cfg.num_key_value_heads // (
+            attn_tp_size * attn_cp_size
+        )
+
+    # tp_size-dependent value (uses mapping.tp_size, not attn_tp_size).
+    assert bindings_cfg.mlp_hidden_size == (cfg.intermediate_size // mapping.tp_size)
+    assert bindings_cfg.tokens_per_block == tokens_per_block