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 15, 22:15 UTC
[ROCm] MI300A compile targets deprecation (#13560)
Gregory Shtrasberg committed
Feb 20, 2025 at 02:05 UTC
0023cd2b9dfe37ee298f9dba8502c5d3c0a1cbbb
4 files changed
+4
-7
CMakeLists.txt
+1
-1
@@ -34,7 +34,7 @@ set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
34
set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
35
36
# Supported AMD GPU architectures.
37
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101")
37
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
38
39
#
40
# Supported/expected torch versions for CUDA/ROCm.
csrc/quantization/fp8/amd/hip_float8_impl.h
+1
-2
@@ -1,7 +1,6 @@
1
#pragma once
2
3
-#if defined(__HIPCC__) && \
4
- (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
3
+#if defined(__HIPCC__) && defined(__gfx942__)
4
#define __HIP__MI300__
5
#endif
6
csrc/rocm/attention.cu
+1
-2
@@ -24,8 +24,7 @@
24
#include "../attention/dtype_fp8.cuh"
25
#include "../quantization/fp8/amd/quant_utils.cuh"
26
27
-#if defined(__HIPCC__) && (defined(__gfx90a__) || defined(__gfx940__) || \
28
- defined(__gfx941__) || defined(__gfx942__))
27
+#if defined(__HIPCC__) && (defined(__gfx90a__) || defined(__gfx942__))
28
#define __HIP__MI300_MI250__
29
#endif
30
vllm/attention/backends/rocm_flash_attn.py
+1
-2
@@ -25,8 +25,7 @@ logger = init_logger(__name__)
25
_PARTITION_SIZE_ROCM = 512
26
_GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
27
_ON_NAVI = "gfx1" in _GPU_ARCH
28
-_ON_MI250_MI300 = any(arch in _GPU_ARCH
29
- for arch in ["gfx90a", "gfx940", "gfx941", "gfx942"])
28
+_ON_MI250_MI300 = any(arch in _GPU_ARCH for arch in ["gfx90a", "gfx942"])
29
30
31
class ROCmFlashAttentionBackend(AttentionBackend):