Accelerate deepseek fp4 b200 ci (#11993)

d98b81e2 · Qiaolin Yu · GitHub · bcecf27e · d98b81e2
Unverified Commit d98b81e2 authored Oct 25, 2025 by Qiaolin Yu Committed by GitHub Oct 25, 2025
Hide whitespace changes
Inline Side-by-side

Showing with 5 additions and 0 deletions

test/srt/test_deepseek_v3_fp4_4gpu.py test/srt/test_deepseek_v3_fp4_4gpu.py +5 -0

No files found.
--- a/test/srt/test_deepseek_v3_fp4_4gpu.py
+++ b/test/srt/test_deepseek_v3_fp4_4gpu.py
@@ -32,6 +32,8 @@ class TestDeepseekV3FP4(CustomTestCase):
            "flashinfer_trtllm",
            "--quantization",
            "modelopt_fp4",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
        ]
        cls.process = popen_launch_server(
            cls.model,
@@ -100,6 +102,9 @@ class TestDeepseekV3FP4MTP(CustomTestCase):
            "1",
            "--speculative-num-draft-tokens",
            "4",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
+            "--enable-beta-spec",
        ]
        cls.process = popen_launch_server(
            cls.model,