[CI] Improve github summary & enable fa3 for more models (#5796)

Author: merrymercy

.github/workflows/pr-test.yml

@@ -123,7 +123,7 @@ jobs:
         timeout-minutes: 10
         run: |
           cd test/srt
-          python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_bs1
+          python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_bs1_default
 
       - name: Benchmark online latency
         timeout-minutes: 10

python/sglang/srt/utils.py

@@ -1970,6 +1970,7 @@ def is_fa3_default_architecture(hf_config):
         "Llama4ForConditionalGeneration",
         "LlamaForCausalLM",
         "MistralForCausalLM",
+        "MixtralForCausalLM",
         "Gemma2ForCausalLM",
         "Gemma3ForConditionalGeneration",
     }

test/srt/run_suite.py

@@ -64,7 +64,7 @@ class TestFile:
         TestFile("test_retract_decode.py", 54),
         TestFile("test_server_args.py", 1),
         TestFile("test_skip_tokenizer_init.py", 117),
-        TestFile("test_srt_engine.py", 237),
+        TestFile("test_srt_engine.py", 261),
         TestFile("test_srt_endpoint.py", 130),
         TestFile("test_torch_compile.py", 76),
         TestFile("test_torch_compile_moe.py", 172),

test/srt/test_bench_one_batch.py

@@ -4,23 +4,22 @@
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_MOE_MODEL_NAME_FOR_TEST,
     CustomTestCase,
-    get_bool_env_var,
     is_in_ci,
     run_bench_one_batch,
     write_github_step_summary,
 )
 
 
 class TestBenchOneBatch(CustomTestCase):
-    def test_bs1(self):
+    def test_bs1_default(self):
         output_throughput = run_bench_one_batch(
             DEFAULT_MODEL_NAME_FOR_TEST, ["--cuda-graph-max-bs", "2"]
         )
 
         if is_in_ci():
             write_github_step_summary(
-                f"### test_bs1\n"
-                f"output_throughput : {output_throughput:.2f} token/s\n"
+                f"### test_bs1_default (llama-3.1-8b)\n"
+                f"output_throughput: {output_throughput:.2f} token/s\n"
             )
             self.assertGreater(output_throughput, 135)
 
@@ -32,9 +31,9 @@ def test_moe_tp2_bs1(self):
         if is_in_ci():
             write_github_step_summary(
                 f"### test_moe_tp2_bs1\n"
-                f"output_throughput : {output_throughput:.2f} token/s\n"
+                f"output_throughput: {output_throughput:.2f} token/s\n"
             )
-            self.assertGreater(output_throughput, 124)
+            self.assertGreater(output_throughput, 125)
 
     def test_torch_compile_tp2_bs1(self):
         output_throughput = run_bench_one_batch(
@@ -45,9 +44,9 @@ def test_torch_compile_tp2_bs1(self):
         if is_in_ci():
             write_github_step_summary(
                 f"### test_torch_compile_tp2_bs1\n"
-                f"output_throughput : {output_throughput:.2f} token/s\n"
+                f"output_throughput: {output_throughput:.2f} token/s\n"
             )
-            self.assertGreater(output_throughput, 225)
+            self.assertGreater(output_throughput, 220)
 
 
 if __name__ == "__main__":

test/srt/test_bench_serving.py

@@ -98,7 +98,7 @@ def test_offline_throughput_with_triton_attention_backend(self):
                 f"### test_offline_throughput_with_triton_attention_backend\n"
                 f'Output throughput: {res["output_throughput"]:.2f} token/s\n'
             )
-            self.assertGreater(res["output_throughput"], 3600)
+            self.assertGreater(res["output_throughput"], 3700)
 
     def test_offline_throughput_default_fp8(self):
         res = run_bench_serving(
@@ -113,7 +113,7 @@ def test_offline_throughput_default_fp8(self):
                 f"### test_offline_throughput_default_fp8\n"
                 f'Output throughput: {res["output_throughput"]:.2f} token/s\n'
             )
-            self.assertGreater(res["output_throughput"], 4200)
+            self.assertGreater(res["output_throughput"], 4300)
 
     def test_online_latency_default(self):
         res = run_bench_serving(
@@ -126,7 +126,7 @@ def test_online_latency_default(self):
         if is_in_ci():
             write_github_step_summary(
                 f"### test_online_latency_default\n"
-                f'median_e2e_latency_ms : {res["median_e2e_latency_ms"]:.2f} ms\n'
+                f'median_e2e_latency_ms: {res["median_e2e_latency_ms"]:.2f} ms\n'
             )
             self.assertLess(res["median_e2e_latency_ms"], 11000)
             self.assertLess(res["median_ttft_ms"], 86)
@@ -161,8 +161,8 @@ def test_online_latency_eagle(self):
         if is_in_ci():
             write_github_step_summary(
                 f"### test_online_latency_eagle\n"
-                f'median_e2e_latency_ms : {res["median_e2e_latency_ms"]:.2f} ms\n'
-                f'accept_length : {res["accept_length"]:.2f} \n'
+                f'median_e2e_latency_ms: {res["median_e2e_latency_ms"]:.2f} ms\n'
+                f'accept_length: {res["accept_length"]:.2f} \n'
             )
             self.assertLess(res["median_e2e_latency_ms"], 900)
             self.assertGreater(res["accept_length"], 3.0)

test/srt/test_full_deepseek_v3.py

@@ -2,7 +2,6 @@
 from types import SimpleNamespace
 
 import requests
-import torch
 
 from sglang.srt.utils import kill_process_tree
 from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
@@ -49,7 +48,7 @@ def test_gsm8k(self):
         metrics = run_eval_few_shot_gsm8k(args)
         print(f"{metrics=}")
 
-        self.assertGreater(metrics["accuracy"], 0.94)
+        self.assertGreater(metrics["accuracy"], 0.935)
 
 
 class TestBenchOneBatch(CustomTestCase):
@@ -58,11 +57,11 @@ def test_bs1(self):
             FULL_DEEPSEEK_V3_MODEL_PATH,
             ["--trust-remote-code", "--tp", "8", "--cuda-graph-max-bs", "2"],
         )
-        print(f"output_throughput : {output_throughput:.2f} token/s")
+        print(f"{output_throughput=:.2f} token/s")
+
         if is_in_ci():
             write_github_step_summary(
-                f"### test_bs1\n"
-                f"output_throughput : {output_throughput:.2f} token/s\n"
+                f"### test_bs1 (deepseek-v3)\n" f"{output_throughput=:.2f} token/s\n"
             )
             self.assertGreater(output_throughput, 70)
 
@@ -121,6 +120,13 @@ def test_gsm8k(self):
         print(f"{avg_spec_accept_length=}")
         self.assertGreater(avg_spec_accept_length, 3.2)
 
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v3)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+                f"{avg_spec_accept_length=:.2f}\n"
+            )
+
 
 if __name__ == "__main__":
     unittest.main()