feat: support deterministic FA4 for GLM-4.7-Flash (#33945)

python/sglang/srt/arg_groups/overrides.py

@@ -2050,6 +2050,7 @@ def _deterministic_is_deepseek_model(view: Any) -> bool:
             "MistralLarge3ForCausalLM",
             "PixtralForConditionalGeneration",
             "GlmMoeDsaForCausalLM",
+            "Glm4MoeLiteForCausalLM",
         ]
     except Exception:
         return False

python/sglang/srt/layers/quantization/unquant.py

@@ -90,11 +90,18 @@ def initialize_bf16_gemm_config(server_args: ServerArgs) -> None:
 
     backend_str = server_args.bf16_gemm_backend
     if backend_str == "auto" and is_sm100_supported():
-        backend_str = "cutedsl"
+        backend_str = (
+            "torch" if server_args.enable_deterministic_inference else "cutedsl"
+        )
 
     backend = Bf16GemmBackend(backend_str)
 
     if backend.is_cutedsl():
+        if server_args.enable_deterministic_inference:
+            raise ValueError(
+                "--bf16-gemm-backend cutedsl is batch-size dependent and cannot "
+                "be combined with --enable-deterministic-inference"
+            )
         if not is_sm100_supported():
             raise ValueError("--bf16-gemm-backend cutedsl requires an SM10x GPU")
 

python/sglang/srt/server_args.py

@@ -1754,7 +1754,7 @@ class ServerArgs:
     bf16_gemm_backend: A[
         str,
         Arg(
-            help="Choose the backend for unquantized BF16 GEMM operations. Options: 'auto' (default; selects 'cutedsl' on SM10x GPUs, otherwise uses cuBLAS via torch.nn.functional.linear), 'cutedsl' (SGLang JIT CuTe DSL TGV BF16 GEMM on SM10x; dispatches between the CuTe DSL kernel and cuBLAS), 'torch' (always uses cuBLAS via torch.nn.functional.linear).",
+            help="Choose the backend for unquantized BF16 GEMM operations. Options: 'auto' (default; selects 'cutedsl' on SM10x GPUs, except deterministic inference selects 'torch'; otherwise uses cuBLAS via torch.nn.functional.linear), 'cutedsl' (SGLang JIT CuTe DSL TGV BF16 GEMM on SM10x; dispatches between the CuTe DSL kernel and cuBLAS), 'torch' (always uses cuBLAS via torch.nn.functional.linear).",
             cli_name="--bf16-gemm-backend",
             choices=BF16_GEMM_BACKEND_CHOICES,
         ),
@@ -8082,6 +8082,7 @@ class ServerArgs:
                         "MistralLarge3ForCausalLM",
                         "PixtralForConditionalGeneration",
                         "GlmMoeDsaForCausalLM",
+                        "Glm4MoeLiteForCausalLM",
                     ]
                 except Exception:
                     pass
@@ -8096,11 +8097,11 @@ class ServerArgs:
                     not in RADIX_SUPPORTED_DETERMINISTIC_ATTENTION_BACKEND
                 ):
                     raise ValueError(
-                        f"Currently only {RADIX_SUPPORTED_DETERMINISTIC_ATTENTION_BACKEND} attention backends are supported for deterministic inference with DeepSeek models. But you're using {attention_backend}."
+                        f"Currently only {RADIX_SUPPORTED_DETERMINISTIC_ATTENTION_BACKEND} attention backends are supported for deterministic inference with absorbed-MLA models. But you're using {attention_backend}."
                     )
                 if attention_backend == "fa4" and not is_sm100_or_sm110_supported():
                     raise ValueError(
-                        "Deterministic inference with DeepSeek models on the fa4 "
+                        "Deterministic inference with absorbed-MLA models on the fa4 "
                         "attention backend requires SM100/SM110: it runs "
                         "absorbed MLA, whose qv argument flash_attn.cute only "
                         "implements on those archs."

python/sglang/test/test_deterministic_utils.py

@@ -31,7 +31,9 @@ class TestDeterministicBase(CustomTestCase):
     def setUpClass(cls):
         cls.model = cls.get_model()
         cls.base_url = DEFAULT_URL_FOR_TEST
-        if "--attention-backend" not in cls.get_server_args():
+        # Identity, not a probe for --attention-backend: a subclass that
+        # deliberately leaves the backend unspecified is a real test case.
+        if cls is TestDeterministicBase:
             raise unittest.SkipTest("Skip the base test class")
 
         cls.process = popen_launch_server(