feat: Add FP8 KV cache support for Triton attention backend (#18882)

test/registered/attention/test_triton_attention_kernels.py

@@ -251,6 +251,8 @@ class TestTritonAttention(CustomTestCase):
             True,
             mask_indptr,
             max_len_extend,
+            1.0,
+            1.0,
         )
 
         b_seq_mask_len = b_seq_len_extend * b_seq_len
@@ -286,6 +288,8 @@ class TestTritonAttention(CustomTestCase):
             True,
             mask_indptr,
             max_len_extend,
+            1.0,
+            1.0,
         )
 
         redundant_attention(
@@ -395,6 +399,8 @@ class TestTritonAttention(CustomTestCase):
             is_causal=True,
             mask_indptr=None,
             max_len_extend=max_len_extend,
+            k_scale=1.0,
+            v_scale=1.0,
             sliding_window_size=WINDOW_SIZE,
         )
 
@@ -517,6 +523,8 @@ class TestTritonAttention(CustomTestCase):
             num_kv_splits,
             max_kv_splits,
             sm_scale,
+            1.0,
+            1.0,
         )
 
         # Correctness reference (float32, stable softmax)
@@ -591,6 +599,7 @@ class TestTritonAttention(CustomTestCase):
             num_kv_splits,
             max_kv_splits,
             sm_scale,
+            1.0,
         )
 
         attn_logits1 = torch.empty(
@@ -616,6 +625,7 @@ class TestTritonAttention(CustomTestCase):
             num_kv_splits,
             max_kv_splits,
             sm_scale,
+            1.0,
         )
 
         cos_sim = torch.nn.functional.cosine_similarity(
@@ -722,6 +732,8 @@ class TestTritonAttention(CustomTestCase):
             is_causal=True,
             mask_indptr=None,
             max_len_extend=max_len_extend,
+            k_scale=1.0,
+            v_scale=1.0,
         )
 
         # Build unified KV indices
@@ -750,6 +762,8 @@ class TestTritonAttention(CustomTestCase):
             o_unified,
             k_buffer,
             v_buffer,
+            1.0,
+            1.0,
             qo_indptr,
             unified_kv_indptr,
             unified_kv_indices,

test/registered/attention/test_wave_attention_kernels.py

@@ -155,6 +155,8 @@ class TestWaveAttention(unittest.TestCase):
             is_causal,
             mask_indptr,
             max_len_extend,
+            1.0,
+            1.0,
         )
 
         o_wave = torch.empty(
@@ -240,6 +242,7 @@ class TestWaveAttention(unittest.TestCase):
             num_kv_splits,
             max_kv_splits,
             sm_scale,
+            1.0,
             logit_cap,
         )
 

test/registered/quant/test_fp8kv_triton.py

@@ -0,0 +1,58 @@
+import unittest
+from types import SimpleNamespace
+from urllib.parse import urlparse
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.few_shot_gsm8k import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_cuda_ci(est_time=520, suite="stage-b-test-large-1-gpu")
+
+
+class TestFP8KVCacheTritonBackend(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "neuralmagic/Meta-Llama-3-8B-Instruct-FP8-KV"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--quantization",
+                "fp8",
+                "--kv-cache-dtype",
+                "fp8_e4m3",
+                "--attention-backend",
+                "triton",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        parsed_url = urlparse(self.base_url)
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=200,
+            host=f"{parsed_url.scheme}://{parsed_url.hostname}",
+            port=parsed_url.port,
+        )
+        metrics = run_eval(args)
+        print(f"{metrics=}")
+        self.assertGreater(metrics["accuracy"], 0.70)
+
+
+if __name__ == "__main__":
+    unittest.main()