A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Compile] Fix AMD Compile Error (#25518)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Sep 23, 2025 at 20:42 UTC
bf68fd76a91bd00f1693a453368688540bf22305
2 files changed
+12
-1
csrc/quantization/activation_kernels.cu
+6
-1
@@ -23,9 +23,14 @@
23
typedef __hip_bfloat162 __nv_bfloat162;
24
typedef __hip_bfloat16 __nv_bfloat16;
25
typedef __hip_bfloat16_raw __nv_bfloat16_raw;
26
-
26
+ #if defined(HIP_FP8_TYPE_OCP)
27
typedef __hip_fp8_e4m3 __nv_fp8_e4m3;
28
typedef __hip_fp8x4_e4m3 __nv_fp8x4_e4m3;
29
+ #else
30
+// ROCm 6.2 fallback: only *_fnuz types exist
31
+typedef __hip_fp8_e4m3_fnuz __nv_fp8_e4m3;
32
+typedef __hip_fp8x4_e4m3_fnuz __nv_fp8x4_e4m3;
33
+ #endif
34
#endif
35
36
#include "core/registration.h"
csrc/rocm/attention.cu
+6
@@ -25,6 +25,12 @@
25
#include "../attention/dtype_fp8.cuh"
26
#include "../quantization/fp8/amd/quant_utils.cuh"
27
28
+// ROCm 6.2 compatibility: map OCP fp8 types to FNUZ variants if OCP is absent
29
+#if !defined(HIP_FP8_TYPE_OCP)
30
+using __hip_fp8_e4m3 = __hip_fp8_e4m3_fnuz;
31
+using __hip_fp8_e5m2 = __hip_fp8_e5m2_fnuz;
32
+#endif
33
+
34
#if defined(__HIPCC__) && \
35
(defined(__gfx90a__) || defined(__gfx942__) || defined(__gfx950__))
36
#define __HIP__GFX9__