[TRTLLM-13948][feat] Set DeepSeekV3 to use Python KV-cache transceiver V2 by default

tensorrt_llm/_torch/models/modeling_deepseekv3.py

@@ -1897,25 +1897,14 @@ class DeepseekV3ForCausalLM(SpecDecOneEngineForCausalLM[DeepseekV3Model,
     def get_preferred_transceiver_runtime(cls,
                                           pretrained_config: Any = None
                                           ) -> Optional[Literal["PYTHON"]]:
-        """GLM-5 family checkpoints default to the Python (v2) KV-cache transceiver.
-
-        This implementation class is shared by DeepSeek-V3/V3.2 and the GLM-5 family — both
-        GLM-5 and GLM-5.2 declare ``GlmMoeDsaForCausalLM`` / ``glm_moe_dsa`` — so the preference
-        is differentiated per checkpoint: only GLM checkpoints opt into the Python transceiver.
-        The MLA backbone transfers a large latent KV, which the Python transceiver handles better
-        in disaggregated serving. This is only adopted when the user leaves
-        ``cache_transceiver_config.transceiver_runtime`` at 'auto' and the effective backend is
-        NIXL; otherwise the C++ transceiver is used.
+        """``DeepseekV3ForCausalLM``, ``DeepseekV32ForCausalLM``, and ``GlmMoeDsaForCausalLM`` all prefer the Python (v2) KV-cache transceiver.
+
+        All three use MLA attention (``DeepseekV3Attention`` and ``DeepseekV32Attention`` both
+        extend ``MLA``), which transfers a large latent KV that the Python transceiver handles
+        better in disaggregated serving. Applied only when ``transceiver_runtime`` is 'auto'
+        and the backend is NIXL.
         """
-        if pretrained_config is None:
-            return None
-        architectures = getattr(pretrained_config, 'architectures', None) or []
-        # model_type is checked as a fallback: it is 'glm_moe_dsa' on GLM
-        # checkpoints until __init__ rewrites it to 'deepseek_v32'.
-        if ("GlmMoeDsaForCausalLM" in architectures or getattr(
-                pretrained_config, 'model_type', None) == 'glm_moe_dsa'):
-            return "PYTHON"
-        return None
+        return "PYTHON"
 
     def __init__(self, model_config: ModelConfig[PretrainedConfig]):
         self.mapping_with_cp = None

tests/scripts/perf-sanity/disaggregated/gb300_deepseek-r1-fp4_128k8k_con256_ctx1_pp4_gen1_dep8_eplb0_mtp1_ccb-NIXL.yaml

@@ -64,6 +64,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 131104
       backend: NIXL
+      transceiver_runtime: CPP
       kv_transfer_timeout_ms: 600000
     disable_overlap_scheduler: true
     speculative_config: &id001
@@ -91,6 +92,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 131104
       backend: NIXL
+      transceiver_runtime: CPP
       kv_transfer_timeout_ms: 600000
     disable_overlap_scheduler: true
     speculative_config: *id001

tests/scripts/perf-sanity/disaggregated/gb300_deepseek-r1-fp4_8k1k_con4096_ctx1_dep4_gen1_dep16_eplb0_mtp1_ccb-NIXL.yaml

@@ -63,6 +63,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: CPP
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -90,5 +91,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: CPP
     disable_overlap_scheduler: true
     speculative_config: *id001

tests/scripts/perf/disaggregated/gb300_deepseek-r1-fp4_128k8k_con256_ctx1_pp4_gen1_dep8_eplb0_mtp1_ccb-NIXL.yaml

@@ -75,6 +75,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 131104
       backend: NIXL
+      transceiver_runtime: CPP
       kv_transfer_timeout_ms: 600000
     disable_overlap_scheduler: true
     speculative_config: &id001
@@ -102,6 +103,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 131104
       backend: NIXL
+      transceiver_runtime: CPP
       kv_transfer_timeout_ms: 600000
     disable_overlap_scheduler: true
     speculative_config: *id001

tests/scripts/perf/disaggregated/gb300_deepseek-r1-fp4_8k1k_con4096_ctx1_dep4_gen1_dep16_eplb0_mtp1_ccb-NIXL.yaml

@@ -74,6 +74,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: CPP
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -101,5 +102,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: CPP
     disable_overlap_scheduler: true
     speculative_config: *id001

tests/unittest/llmapi/test_llm_args.py

@@ -3760,13 +3760,8 @@ def test_resolve_default_backend_env_priority(self, monkeypatch):
             backend="UCX")._resolve_default_backend() == ("UCX", None)
 
 
-class TestGlm5TransceiverPreference:
-    """GLM-5 defaults to the Python KV-cache transceiver in disagg.
-
-    DeepseekV3ForCausalLM is shared by DeepSeek-V3/V3.2 and GLM-5
-    (GlmMoeDsaForCausalLM); the preference must apply to GLM checkpoints
-    only.
-    """
+class TestDeepseekTransceiverPreference:
+    """DeepseekV3ForCausalLM, DeepseekV32ForCausalLM, and GlmMoeDsaForCausalLM all prefer the Python KV-cache transceiver."""
 
     @staticmethod
     def _pretrained_config(architectures, model_type):
@@ -3775,43 +3770,34 @@ def _pretrained_config(architectures, model_type):
         cfg.model_type = model_type
         return cfg
 
-    @pytest.mark.parametrize(
-        "architectures,model_type,expected",
-        [
-            (["GlmMoeDsaForCausalLM"], "glm_moe_dsa", "PYTHON"),
-            (["DeepseekV3ForCausalLM"], "deepseek_v3", None),
-            (["DeepseekV32ForCausalLM"], "deepseek_v32", None),
-            # Each predicate in isolation: the architecture match and the
-            # model_type fallback must each suffice on their own.
-            (["GlmMoeDsaForCausalLM"], "deepseek_v32", "PYTHON"),
-            (["DeepseekV32ForCausalLM"], "glm_moe_dsa", "PYTHON"),
-        ])
-    def test_preference_per_architecture(self, architectures, model_type,
-                                         expected):
+    @pytest.mark.parametrize("architectures,model_type", [
+        (["GlmMoeDsaForCausalLM"], "glm_moe_dsa"),
+        (["DeepseekV3ForCausalLM"], "deepseek_v3"),
+        (["DeepseekV32ForCausalLM"], "deepseek_v32"),
+    ])
+    def test_preference_per_architecture(self, architectures: list[str],
+                                         model_type: str) -> None:
         from tensorrt_llm._torch.models.modeling_deepseekv3 import \
             DeepseekV3ForCausalLM
         cfg = self._pretrained_config(architectures, model_type)
         assert DeepseekV3ForCausalLM.get_preferred_transceiver_runtime(
-            cfg) == expected
+            cfg) == "PYTHON"
 
-    def test_no_config_defers_to_cpp(self):
-        """Without a pretrained config the class defers to the C++ default."""
+    def test_prefers_python_without_config(self) -> None:
+        """Preference is unconditional without a pretrained config."""
         from tensorrt_llm._torch.models.modeling_deepseekv3 import \
             DeepseekV3ForCausalLM
-        assert DeepseekV3ForCausalLM.get_preferred_transceiver_runtime() is None
-
-    def test_glm5_resolves_auto_to_python_on_nixl(self):
-        """GLM-5 on NIXL adopts the Python transceiver from 'auto'.
+        assert DeepseekV3ForCausalLM.get_preferred_transceiver_runtime(
+        ) == "PYTHON"
 
-        End-to-end through _resolve_transceiver_runtime_auto with the real
-        model class and a GLM pretrained config.
-        """
+    def test_deepseek_resolves_auto_to_python_on_nixl(self) -> None:
+        """DeepseekV3ForCausalLM on NIXL adopts the Python transceiver from 'auto'."""
         from tensorrt_llm._torch.models.modeling_deepseekv3 import \
             DeepseekV3ForCausalLM
         args = TorchLlmArgs(
             model="/tmp/dummy_model",
             cache_transceiver_config=CacheTransceiverConfig(backend="NIXL"),
         )
-        cfg = self._pretrained_config(["GlmMoeDsaForCausalLM"], "glm_moe_dsa")
+        cfg = self._pretrained_config(["DeepseekV3ForCausalLM"], "deepseek_v3")
         _resolve_transceiver_runtime_auto(args, DeepseekV3ForCausalLM, cfg)
         assert args.cache_transceiver_config.transceiver_runtime == "PYTHON"