[AMD] Fix GLM-5 fp8 KV quant path dispatch on MI300 (#22314)

python/sglang/srt/mem_cache/memory_pool.py

@@ -45,7 +45,7 @@ from sglang.srt.layers.attention.nsa.quant_k_cache import (
     quantize_k_cache,
     quantize_k_cache_separate,
 )
-from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
+from sglang.srt.layers.quantization.fp8_kernel import fp8_dtype, is_fp8_fnuz
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.mem_cache.utils import (
     get_mla_kv_buffer_triton,
@@ -1575,13 +1575,9 @@ class MLATokenToKVPool(KVCache):
     ):
         layer_id = layer.layer_id
 
-        if self.nsa_kv_cache_store_fp8:
+        if _is_hip and self.use_nsa and self.dtype == fp8_dtype:
-            if _is_hip:
             # HIP FP8 path uses raw MLA KV layout (nope + rope) without per-block scales.
             # Fuse BF16/FP16 -> FP8 cast with paged KV write.
-                fp8_dtype = (
-                    torch.float8_e4m3fnuz if _is_fp8_fnuz else torch.float8_e4m3fn
-                )
             set_mla_kv_buffer_triton_fp8_quant(
                 self.kv_buffer[layer_id - self.start_layer],
                 loc,
@@ -1589,7 +1585,7 @@ class MLATokenToKVPool(KVCache):
                 cache_k_rope,
                 fp8_dtype,
             )
-            else:
+        elif self.nsa_kv_cache_store_fp8:
             # OPTIMIZATION: Quantize k_nope and k_rope separately to avoid concat overhead
             # This also enables reuse of set_mla_kv_buffer_triton two-tensor write path
             # quantize_k_cache_separate returns (nope_part, rope_part) as uint8 bytes