Remove mi300a (ROCm#440)

* Removing gfx940 and gfx941 targets. These have been deprecated in favor of gfx942 for MI300X

Signed-off-by: Gregory Shtrasberg <[email protected]>

* Remove from custom kernels as well

---------

Signed-off-by: Gregory Shtrasberg <[email protected]>

Author: gshtras

CMakeLists.txt

@@ -34,7 +34,7 @@ set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.

csrc/quantization/fp8/amd/hip_float8_impl.h

@@ -1,7 +1,6 @@
 #pragma once
 
-#if defined(__HIPCC__) && \
-    (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+#if defined(__HIPCC__) && defined(__gfx942__)
   #define __HIP__MI300__
 #endif
 

csrc/rocm/attention.cu

@@ -24,8 +24,7 @@
 #include "../attention/dtype_fp8.cuh"
 #include "../quantization/fp8/amd/quant_utils.cuh"
 
-#if defined(__HIPCC__) && (defined(__gfx90a__) || defined(__gfx940__) || \
-                           defined(__gfx941__) || defined(__gfx942__))
+#if defined(__HIPCC__) && (defined(__gfx90a__) || defined(__gfx942__))
   #define __HIP__MI300_MI250__
 #endif
 

csrc/rocm/custom_kernels.cu

@@ -5,13 +5,11 @@
 #include <algorithm>
 #include "cuda_compat.h"
 
-#if defined(__HIPCC__) && (defined(__gfx90a__) || defined(__gfx940__) || \
-                           defined(__gfx941__) || defined(__gfx942__))
+#if defined(__HIPCC__) && (defined(__gfx90a__) || defined(__gfx942__))
   #define __HIP__MI300_MI250__
 #endif
 
-#if defined(__HIPCC__) && \
-    (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+#if defined(__HIPCC__) && defined(__gfx942__)
   #define __HIP__MI300__
 #endif
 

vllm/attention/backends/rocm_flash_attn.py

@@ -25,8 +25,7 @@
 _PARTITION_SIZE_ROCM = 256
 _GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
 _ON_NAVI = "gfx1" in _GPU_ARCH
-_ON_MI250_MI300 = any(arch in _GPU_ARCH
-                      for arch in ["gfx90a", "gfx940", "gfx941", "gfx942"])
+_ON_MI250_MI300 = any(arch in _GPU_ARCH for arch in ["gfx90a", "gfx942"])
 
 
 class ROCmFlashAttentionBackend(AttentionBackend):