[AMD] deepseek-v4 clean env vars (#28920)

python/sglang/srt/layers/attention/deepseek_v4_backend_hip_radix.py

@@ -1472,13 +1472,11 @@ class DeepseekV4HipRadixBackend(
                     extra_indices.shape[-1] % 64 == 0
                 ), f"{extra_indices.shape=}'s last dimension is not aligned to 64"
 
-            import os
-
             from sglang.srt.layers.attention.hip_flash_mla import (
                 flash_mla_with_kvcache_entrypoint,
             )
 
-            backend = os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "kernel")
+            backend = envs.SGLANG_HACK_FLASHMLA_BACKEND.get()
             input_dict = dict(
                 q=q,
                 k_cache=swa_k_cache,

python/sglang/srt/layers/attention/dsv4/unified_kv_kernels/env_gate.py

@@ -1,15 +1,12 @@
 from __future__ import annotations
 
 import functools
-import os
 
+from sglang.srt.environ import envs
 from sglang.srt.utils import is_hip
 
 
 @functools.lru_cache(maxsize=1)
 def is_unified_kv_triton() -> bool:
     # unified_kv_triton is only implemented on HIP (ROCm)
-    return (
-        is_hip()
-        and os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "") == "unified_kv_triton"
-    )
+    return is_hip() and envs.SGLANG_HACK_FLASHMLA_BACKEND.get() == "unified_kv_triton"

python/sglang/srt/layers/attention/hip_flash_mla.py

@@ -2,6 +2,7 @@ from typing import Any, Optional
 
 import torch
 
+from sglang.srt.environ import envs
 from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
 from sglang.srt.utils import is_hip
 
@@ -10,9 +11,7 @@ FP8_DTYPE = torch.float8_e4m3fnuz if is_fp8_fnuz() else torch.float8_e4m3fn
 
 def flash_mla_with_kvcache_entrypoint(backend: str, **kwargs):
     if is_hip():
-        import os
-
-        backend = os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "tilelang")
+        backend = envs.SGLANG_HACK_FLASHMLA_BACKEND.get()
     else:
         import sgl_kernel.flash_mla as flash_mla
 

python/sglang/srt/server_args.py

@@ -2626,6 +2626,18 @@ class ServerArgs:
                 envs.SGLANG_OPT_USE_TILELANG_MHC_PRE.set(False)
                 envs.SGLANG_OPT_DEEPGEMM_HC_PRENORM.set(False)
                 envs.SGLANG_FP8_PAGED_MQA_LOGITS_TORCH.set(True)
+            elif is_hip():
+                envs.SGLANG_OPT_DEEPGEMM_HC_PRENORM.set(False)
+                envs.SGLANG_OPT_USE_FUSED_COMPRESS.set(True)
+                envs.SGLANG_OPT_FP8_WO_A_GEMM.set(False)
+                envs.SGLANG_OPT_USE_JIT_INDEXER_METADATA.set(False)
+                envs.SGLANG_OPT_USE_TOPK_V2.set(False)
+                envs.SGLANG_OPT_USE_AITER_INDEXER.set(True)
+                envs.SGLANG_OPT_USE_TILELANG_MHC_PRE.set(False)
+                envs.SGLANG_OPT_USE_TILELANG_MHC_POST.set(False)
+                envs.SGLANG_FP8_PAGED_MQA_LOGITS_TORCH.set(True)
+                envs.SGLANG_OPT_USE_MULTI_STREAM_OVERLAP.set(False)
+                envs.SGLANG_EAGER_INPUT_NO_COPY.set(True)
 
         elif model_arch in ["GptOssForCausalLM"]:
             # Set attention backend for GPT-OSS