Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

Add gfx1200 & gfx1201 support on linux (#9878)

saman-amdead27aa9fe85
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 034fc7d79..e2447f32a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -86,9 +86,9 @@ if(CMAKE_CUDA_COMPILER)
)
endif()
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a):xnack[+-]$"
+set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a|1200|1201):xnack[+-]$"
CACHE STRING
- "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a):xnack[+-]$\"."
+ "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a|1200|1201):xnack[+-]$\"."
)
check_language(HIP)
@@ -97,7 +97,7 @@ if(CMAKE_HIP_COMPILER)
find_package(hip REQUIRED)
if(NOT AMDGPU_TARGETS)
- list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012])$")
+ list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012]|120[01])$")
elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
endif()
diff --git a/CMakePresets.json b/CMakePresets.json
index f3507e3f2..bf99917f0 100644
--- a/CMakePresets.json
+++ b/CMakePresets.json
@@ -56,7 +56,7 @@
"name": "ROCm 6",
"inherits": [ "ROCm" ],
"cacheVariables": {
- "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
+ "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
}
}
],
diff --git a/llama/patches/0022-add-rdna4-support.patch b/llama/patches/0022-add-rdna4-support.patch
new file mode 100644
index 000000000..9b77ae84c
--- /dev/null
+++ b/llama/patches/0022-add-rdna4-support.patch
@@ -0,0 +1,103 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Saman <[email protected]>
+Date: Wed, 19 Mar 2025 14:02:26 -0700
+Subject: [PATCH] add rdna4 support
+
+---
+ ggml/src/ggml-cuda/common.cuh | 6 ++++--
+ ggml/src/ggml-cuda/mmq.cu | 2 +-
+ ggml/src/ggml-cuda/mmq.cuh | 4 ++--
+ ggml/src/ggml-cuda/mmvq.cu | 4 ++--
+ ggml/src/ggml-cuda/vendors/hip.h | 4 ++++
+ 5 files changed, 13 insertions(+), 7 deletions(-)
+
+diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
+index adf0d3ec..b24593fc 100644
+--- a/ggml/src/ggml-cuda/common.cuh
++++ b/ggml/src/ggml-cuda/common.cuh
+@@ -61,11 +61,13 @@
+ #define GGML_CUDA_CC_RDNA1 (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
+ #define GGML_CUDA_CC_RDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
+ #define GGML_CUDA_CC_RDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
++#define GGML_CUDA_CC_RDNA4 (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
+
+ #define GGML_CUDA_CC_IS_RDNA(cc) (cc >= GGML_CUDA_CC_RDNA1)
+ #define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
+ #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
+-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
++#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
++#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
+ #define GGML_CUDA_CC_IS_GCN(cc) (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
+ #define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
+
+@@ -386,7 +388,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
+ #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+ #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
+ c = __builtin_amdgcn_sdot4(a, b, c, false);
+-#elif defined(RDNA3)
++#elif defined(RDNA3) || defined(RDNA4)
+ c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
+ #elif defined(__gfx1010__) || defined(__gfx900__)
+ int tmp1;
+diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
+index 10f2ebb1..933d945c 100644
+--- a/ggml/src/ggml-cuda/mmq.cu
++++ b/ggml/src/ggml-cuda/mmq.cu
+@@ -149,5 +149,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
+ return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+ }
+
+- return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
++ return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+ }
+diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
+index 0451c65f..66ce2bc9 100644
+--- a/ggml/src/ggml-cuda/mmq.cuh
++++ b/ggml/src/ggml-cuda/mmq.cuh
+@@ -2577,9 +2577,9 @@ static __device__ void mul_mat_q_process_tile(
+
+ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
+ #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+-#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
++#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+ __launch_bounds__(WARP_SIZE*nwarps, 2)
+-#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
++#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+ #else
+ #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+ __launch_bounds__(WARP_SIZE*nwarps, 1)
+diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
+index 4fb466ca..23ae7abc 100644
+--- a/ggml/src/ggml-cuda/mmvq.cu
++++ b/ggml/src/ggml-cuda/mmvq.cu
+@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
+
+ constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+
+-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
++#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4))
+ constexpr int nwarps = 1;
+ constexpr int rows_per_cuda_block = 1;
+ #else
+ constexpr int nwarps = ncols_y <= 4 ? 4 : 2;
+ constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
+-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
++#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) && !defined(RDNA4)
+
+ const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+ const int row0 = rows_per_cuda_block*blockIdx.x;
+diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
+index 81964611..a62544b5 100644
+--- a/ggml/src/ggml-cuda/vendors/hip.h
++++ b/ggml/src/ggml-cuda/vendors/hip.h
+@@ -150,6 +150,10 @@
+ #define CDNA
+ #endif
+
++#if defined(__gfx1200__) || defined(__gfx1201__)
++#define RDNA4
++#endif
++
+ #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+ defined(__gfx1150__) || defined(__gfx1151__)
+ #define RDNA3
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/common.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/common.cuh
index adf0d3ecb..b24593fc2 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/common.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/common.cuh
@@ -61,11 +61,13 @@
#define GGML_CUDA_CC_RDNA1 (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
#define GGML_CUDA_CC_RDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
#define GGML_CUDA_CC_RDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
+#define GGML_CUDA_CC_RDNA4 (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
#define GGML_CUDA_CC_IS_RDNA(cc) (cc >= GGML_CUDA_CC_RDNA1)
#define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
#define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
+#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
#define GGML_CUDA_CC_IS_GCN(cc) (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
#define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
@@ -386,7 +388,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(RDNA3)
+#elif defined(RDNA3) || defined(RDNA4)
c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
#elif defined(__gfx1010__) || defined(__gfx900__)
int tmp1;
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu b/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu
index 10f2ebb1c..933d945c8 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu
@@ -149,5 +149,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
}
- return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+ return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
}
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh
index 0451c65f3..66ce2bc95 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh
@@ -2577,9 +2577,9 @@ static __device__ void mul_mat_q_process_tile(
template <ggml_type type, int mmq_x, int nwarps, bool need_check>
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
__launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
#else
#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
__launch_bounds__(WARP_SIZE*nwarps, 1)
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu b/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu
index 4fb466ca0..23ae7abcb 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu
@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4))
constexpr int nwarps = 1;
constexpr int rows_per_cuda_block = 1;
#else
constexpr int nwarps = ncols_y <= 4 ? 4 : 2;
constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) && !defined(RDNA4)
const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
const int row0 = rows_per_cuda_block*blockIdx.x;
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h b/ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h
index 81964611c..a62544b5a 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h
@@ -150,6 +150,10 @@
#define CDNA
#endif
+#if defined(__gfx1200__) || defined(__gfx1201__)
+#define RDNA4
+#endif
+
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
defined(__gfx1150__) || defined(__gfx1151__)
#define RDNA3