[AMD] feat: add DLLM support for AMD GPUs with LLaDA2 testing (#15560)

.github/workflows/pr-test-amd.yml

@@ -796,6 +796,43 @@ jobs:
         run: |
           bash scripts/ci/amd_ci_exec.sh -e SGLANG_USE_AITER_AR=0 -e SGLANG_USE_AITER=0 -e HF_HUB_ENABLE_HF_TRANSFER=0 python3 test_moe_eval_accuracy_large.py
 
+  dllm-test-1-gpu-amd:
+    needs: [check-changes, stage-a-test-1-amd]
+    if: |
+      always() &&
+      (
+        (inputs.target_stage == 'dllm-test-1-gpu-amd') ||
+        (
+          !inputs.target_stage &&
+          (!failure() && !cancelled()) &&
+          ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
+        )
+      )
+    strategy:
+      fail-fast: false
+      matrix:
+        runner: [linux-mi325-gpu-1]
+    runs-on: ${{matrix.runner}}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ensure_vram_clear.sh rocm
+
+      - name: Start CI container
+        run: bash scripts/ci/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd_ci_install_dependency.sh
+
+      - name: Test DLLM (LLaDA2)
+        timeout-minutes: 30
+        run: |
+          bash scripts/ci/amd_ci_exec.sh python3 dllm/test_llada2_mini_amd.py
+
   pr-test-amd-finish:
     needs:
       [
@@ -814,7 +851,8 @@ jobs:
         performance-test-1-gpu-part-2-amd,
         performance-test-2-gpu-amd,
         accuracy-test-1-gpu-amd,
-        accuracy-test-2-gpu-amd, # Temporarily disabled
+        accuracy-test-2-gpu-amd,
+        dllm-test-1-gpu-amd,
       ]
     if: always()
     runs-on: ubuntu-latest

python/sglang/srt/model_executor/forward_batch_info.py

@@ -53,7 +53,7 @@ from sglang.srt.layers.dp_attention import (
     set_is_extend_in_batch,
 )
 from sglang.srt.server_args import get_global_server_args
-from sglang.srt.utils import get_compiler_backend, is_npu, support_triton
+from sglang.srt.utils import get_compiler_backend, is_hip, is_npu, support_triton
 from sglang.srt.utils.common import ceil_align
 
 if TYPE_CHECKING:
@@ -490,13 +490,15 @@ class ForwardBatch:
         # Override the positions with diffusion LLM or spec_info
         if batch.dllm_config is not None:
             block_size = batch.dllm_config.block_size
+            # Use int64 for AMD rotary embedding kernel compatibility
+            positions_dtype = torch.int64 if is_hip() else torch.int32
             ret.positions = torch.tensor(
                 [
                     i
                     for block_offset in batch.dllm_block_offsets
                     for i in range(block_offset, block_offset + block_size)
                 ],
-                dtype=torch.int32,
+                dtype=positions_dtype,
             ).to(device, non_blocking=True)
         elif (
             ret.spec_info is not None

python/sglang/srt/server_args.py

@@ -2425,7 +2425,19 @@ class ServerArgs:
     def _handle_dllm_inference(self):
         if self.dllm_algorithm is None:
             return
-        if not self.disable_cuda_graph:
+        # On AMD/HIP, disable cuda graph for DLLM and use triton backend
+        if is_hip():
+            if not self.disable_cuda_graph:
+                logger.warning(
+                    "Cuda graph is disabled for diffusion LLM inference on AMD GPUs"
+                )
+                self.disable_cuda_graph = True
+            if self.attention_backend not in ["triton", "aiter"]:
+                logger.warning(
+                    "Attention backend is set to triton for diffusion LLM inference on AMD GPUs"
+                )
+                self.attention_backend = "triton"
+        elif not self.disable_cuda_graph:
             if self.cuda_graph_bs != [1]:
                 logger.warning(
                     "Cuda graph bs is set to [1] because of using diffusion LLM inference"

test/srt/dllm/test_llada2_mini_amd.py

@@ -0,0 +1,87 @@
+"""
+Test LLaDA2 (Diffusion Language Model) on AMD GPUs.
+
+This test verifies that DLLM works on AMD with triton attention backend.
+"""
+
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+
+class TestLLaDA2MiniAMD(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "inclusionAI/LLaDA2.0-mini"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+
+        other_args = [
+            "--trust-remote-code",
+            "--mem-fraction-static",
+            "0.9",
+            "--max-running-requests",
+            "1",
+            "--attention-backend",
+            "triton",  # Use triton for AMD instead of flashinfer
+            "--dllm-algorithm",
+            "LowConfidence",
+        ]
+
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        """Test GSM8K accuracy with DLLM on AMD."""
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        # Relaxed thresholds for AMD - may need adjustment
+        self.assertGreater(metrics["accuracy"], 0.80)
+        self.assertGreater(metrics["output_throughput"], 50)
+
+    def test_bs_1_speed(self):
+        """Test single batch inference speed."""
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (llada2-mini AMD) with tp1\n"
+                f"{speed=:.2f} token/s\n"
+            )
+            # Relaxed threshold for AMD
+            self.assertGreater(speed, 10)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/run_suite.py

@@ -203,6 +203,7 @@ suite_amd = {
         # TestFile("lora/test_lora_backend.py", 99), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
         # TestFile("lora/test_lora_cuda_graph.py", 250), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
         # TestFile("lora/test_lora_qwen3.py", 97), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
+        TestFile("dllm/test_llada2_mini_amd.py", 520),
         TestFile("models/test_compressed_tensors_models.py", 42),
         TestFile("models/test_cross_encoder_models.py", 150),
         TestFile("models/test_qwen_models.py", 82),