From b1c7ae7f0fecd3be5bfd2efbe2215cb2932a92f7 Mon Sep 17 00:00:00 2001
From: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
Date: Fri, 17 Jul 2026 03:27:16 +0000
Subject: [PATCH 1/2] [None][feat] Default GLM-5 family to the Python KV-cache
 transceiver in disagg

GLM-5 and GLM-5.2 (GlmMoeDsaForCausalLM / glm_moe_dsa) share the
DeepseekV3ForCausalLM implementation class; override
get_preferred_transceiver_runtime() to opt GLM checkpoints into the
Python (v2) transceiver, differentiated per checkpoint so DeepSeek-V3/
V3.2 keep the C++ default. The preference is adopted only when the user
leaves cache_transceiver_config.transceiver_runtime at 'auto' and the
effective backend is NIXL.

Set transceiver_runtime: PYTHON explicitly in the GLM-5 perf-sanity
disaggregated NIXL configs, and switch the GLM-5 disagg stress-test
config from backend DEFAULT to NIXL + PYTHON so the
TRTLLM_USE_UCX_KVCACHE=1 fallback cannot silently revert it to the C++
transceiver.

Add unit tests covering the per-architecture preference and the
end-to-end 'auto' resolution for GLM-5 on NIXL.

Signed-off-by: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
---
 .../_torch/models/modeling_deepseekv3.py      | 29 ++++++++++-
 ...txtp4ep4_gentp4ep4_glm5_nvfp4_dp_tllm.yaml |  6 ++-
 ...x1_dep4_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml |  2 +
 ..._dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml |  2 +
 ...1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml |  2 +
 ..._dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml |  2 +
 ...x1_dep4_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml |  2 +
 ...1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml |  2 +
 ...x1_dep2_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml |  2 +
 ..._dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml |  2 +
 ...1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml |  2 +
 ..._dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml |  2 +
 ...x1_dep2_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml |  2 +
 ...1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml |  2 +
 tests/unittest/llmapi/test_llm_args.py        | 52 +++++++++++++++++++
 15 files changed, 108 insertions(+), 3 deletions(-)

diff --git a/tensorrt_llm/_torch/models/modeling_deepseekv3.py b/tensorrt_llm/_torch/models/modeling_deepseekv3.py
index 2ccc34a3cda4..f7933f8653d2 100755
--- a/tensorrt_llm/_torch/models/modeling_deepseekv3.py
+++ b/tensorrt_llm/_torch/models/modeling_deepseekv3.py
@@ -28,7 +28,7 @@
 import copy
 import math
 import os
-from typing import Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Literal, Optional, Tuple
 
 import torch
 import triton
@@ -1893,6 +1893,33 @@ def forward(
 class DeepseekV3ForCausalLM(SpecDecOneEngineForCausalLM[DeepseekV3Model,
                                                         PretrainedConfig]):
 
+    @classmethod
+    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.
+        """
+        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
+
     def __init__(self, model_config: ModelConfig[PretrainedConfig]):
         self.mapping_with_cp = None
         # Note: Currently the usage of mapping is all over the place making its usage brittle
diff --git a/tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp4ep4_gentp4ep4_glm5_nvfp4_dp_tllm.yaml b/tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp4ep4_gentp4ep4_glm5_nvfp4_dp_tllm.yaml
index 57cf5bcbebed..fc790618269b 100644
--- a/tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp4ep4_gentp4ep4_glm5_nvfp4_dp_tllm.yaml
+++ b/tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp4ep4_gentp4ep4_glm5_nvfp4_dp_tllm.yaml
@@ -20,7 +20,8 @@ context_servers:
   cuda_graph_config: null
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384
 generation_servers:
   num_instances: 1
@@ -55,5 +56,6 @@ generation_servers:
     - 1024
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384
diff --git a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con1_ctx1_dep4_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con1_ctx1_dep4_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml
index 6678fdfc4f8b..9a514f62dc65 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con1_ctx1_dep4_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con1_ctx1_dep4_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml
@@ -61,6 +61,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -89,5 +90,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con4096_ctx1_dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con4096_ctx1_dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
index a54e7d0335cf..4b01710438c2 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con4096_ctx1_dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con4096_ctx1_dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
@@ -66,6 +66,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -93,5 +94,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con512_ctx1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con512_ctx1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
index e8bcff454293..1b427f4f2c05 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con512_ctx1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_1k1k_con512_ctx1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
@@ -63,6 +63,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     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: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con1024_ctx1_dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con1024_ctx1_dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
index 37d162efda13..9dc48b40047f 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con1024_ctx1_dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con1024_ctx1_dep4_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
@@ -66,6 +66,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -93,5 +94,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con1_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con1_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml
index 3761761e4e9d..d9fef4a05711 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con1_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con1_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml
@@ -61,6 +61,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -89,5 +90,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con512_ctx1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con512_ctx1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
index 876121e49a57..6f95b245621a 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con512_ctx1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb200_glm-5-fp4_8k1k_con512_ctx1_dep4_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
@@ -63,6 +63,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     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: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con1_ctx1_dep2_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con1_ctx1_dep2_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml
index 7e5aa487facc..e1f98578eda1 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con1_ctx1_dep2_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con1_ctx1_dep2_gen1_tep4_eplb0_mtp3_ccb-NIXL.yaml
@@ -61,6 +61,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -89,5 +90,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con4096_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con4096_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
index f6334f686d1c..e2d50f587b7b 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con4096_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con4096_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
@@ -66,6 +66,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -93,5 +94,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
index 8ef0b3db0473..ebc92211e8ee 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
@@ -63,6 +63,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     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: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1024_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1024_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
index 22a64298673b..0260accf37ea 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1024_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1024_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml
@@ -66,6 +66,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -93,5 +94,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1_ctx1_dep2_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1_ctx1_dep2_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml
index 1e6c43b2ddf7..67ae7c9d5c3a 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1_ctx1_dep2_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1_ctx1_dep2_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml
@@ -61,6 +61,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: &id001
       decoding_type: MTP
@@ -89,5 +90,6 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
index 777987a01492..8a8ddd2f7e96 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml
@@ -63,6 +63,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 16384
       backend: NIXL
+      transceiver_runtime: PYTHON
     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: PYTHON
     disable_overlap_scheduler: true
     speculative_config: *id001
diff --git a/tests/unittest/llmapi/test_llm_args.py b/tests/unittest/llmapi/test_llm_args.py
index 4f17a87123c3..75de54aebb5c 100644
--- a/tests/unittest/llmapi/test_llm_args.py
+++ b/tests/unittest/llmapi/test_llm_args.py
@@ -3338,3 +3338,55 @@ def test_resolve_default_backend_env_priority(self, monkeypatch):
         # An explicit backend bypasses the env vars entirely.
         assert CacheTransceiverConfig(
             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.
+    """
+
+    @staticmethod
+    def _pretrained_config(architectures, model_type):
+        from unittest.mock import MagicMock
+        cfg = MagicMock()
+        cfg.architectures = architectures
+        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),
+    ])
+    def test_preference_per_architecture(self, architectures, model_type,
+                                         expected):
+        from tensorrt_llm._torch.models.modeling_deepseekv3 import \
+            DeepseekV3ForCausalLM
+        cfg = self._pretrained_config(architectures, model_type)
+        assert DeepseekV3ForCausalLM.get_preferred_transceiver_runtime(
+            cfg) == expected
+
+    def test_no_config_defers_to_cpp(self):
+        """Without a pretrained config the class defers to the C++ default."""
+        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'.
+
+        End-to-end through _resolve_transceiver_runtime_auto with the real
+        model class and a GLM pretrained config.
+        """
+        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")
+        _resolve_transceiver_runtime_auto(args, DeepseekV3ForCausalLM, cfg)
+        assert args.cache_transceiver_config.transceiver_runtime == "PYTHON"

From 49fe11766c44a0700da723872f69f0b8246d2b9b Mon Sep 17 00:00:00 2001
From: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
Date: Fri, 17 Jul 2026 05:56:55 +0000
Subject: [PATCH 2/2] [None][test] Isolate architecture and model_type
 predicates in GLM transceiver test

Address review: the GLM-positive case satisfied both predicates at
once, so deleting either production check still passed. Add one case
per predicate in isolation (GLM architecture with a non-GLM model_type,
and a non-GLM architecture with model_type=glm_moe_dsa).

Signed-off-by: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
---
 tests/unittest/llmapi/test_llm_args.py | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/tests/unittest/llmapi/test_llm_args.py b/tests/unittest/llmapi/test_llm_args.py
index 75de54aebb5c..1060bbe0ce8d 100644
--- a/tests/unittest/llmapi/test_llm_args.py
+++ b/tests/unittest/llmapi/test_llm_args.py
@@ -3356,11 +3356,17 @@ 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),
-    ])
+    @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):
         from tensorrt_llm._torch.models.modeling_deepseekv3 import \