From d7fad6100a32a369ba9bd0e420d9471d71ec8ade Mon Sep 17 00:00:00 2001 From: Ilya Panfilov Date: Tue, 27 Oct 2026 01:07:08 -0400 Subject: [PATCH 1/3] Use gfx1250-strict AITER and CK. Add gfx1250-strict AITER V3 filter --- 3rdparty/QoLA | 2 +- 3rdparty/ck_jit | 2 +- transformer_engine/common/fused_attn_rocm/fused_attn_ck.cpp | 6 ++++-- 3 files changed, 6 insertions(+), 4 deletions(-) diff --git a/3rdparty/QoLA b/3rdparty/QoLA index a2bfad643b..c33917c7e0 160000 --- a/3rdparty/QoLA +++ b/3rdparty/QoLA @@ -1 +1 @@ -Subproject commit a2bfad643b5a4b8f7e6447063a713a5f193bfb65 +Subproject commit c33917c7e01d300615f7ef8ed355082a144bbed4 diff --git a/3rdparty/ck_jit b/3rdparty/ck_jit index 2f1859c434..c0392304ae 160000 --- a/3rdparty/ck_jit +++ b/3rdparty/ck_jit @@ -1 +1 @@ -Subproject commit 2f1859c434bfa2a1d83004ed424ec12cef595126 +Subproject commit c0392304ae352e1cbb40f5ec0cd43ba46c5bdac2 diff --git a/transformer_engine/common/fused_attn_rocm/fused_attn_ck.cpp b/transformer_engine/common/fused_attn_rocm/fused_attn_ck.cpp index c90e60fa4f..86ca68203d 100644 --- a/transformer_engine/common/fused_attn_rocm/fused_attn_ck.cpp +++ b/transformer_engine/common/fused_attn_rocm/fused_attn_ck.cpp @@ -499,7 +499,8 @@ void fused_attn_ck_fwd_impl( const bool nvte_log_ck_config = getenv("NVTE_LOG_CK_CONFIG"); // ASM v3 does not support sink; force CK tile path - bool nvte_ck_uses_fwd_v3 = !has_sink && getenv("NVTE_CK_USES_FWD_V3", 1); + bool nvte_ck_uses_fwd_v3 = !has_sink && getenv("NVTE_CK_USES_FWD_V3", 1) && + (cuda::sm_arch() != 125 || !cuda::sm_arch_name().starts_with("gfx1250-strict")); int nvte_ck_how_v3_bf16_cvt = getenv("NVTE_CK_HOW_V3_BF16_CVT", 1); bool nvte_ck_zero_out_pad = getenv("NVTE_CK_ZERO_OUT_PAD", 1); NVTE_QKV_Format qkv_format = nvte_get_qkv_format(layout); @@ -856,7 +857,8 @@ void fused_attn_ck_bwd_impl( // ASM v3 bwd does not compute the sink gradient; force the CK tile path so // has_sink users still get a correct (if slower) d_sink, mirroring the fwd // ASM v3 sink guard above. - bool nvte_ck_uses_bwd_v3 = getenv("NVTE_CK_USES_BWD_V3", 1) && !has_sink; + bool nvte_ck_uses_bwd_v3 = !has_sink && getenv("NVTE_CK_USES_BWD_V3", 1) && + (cuda::sm_arch() != 125 || !cuda::sm_arch_name().starts_with("gfx1250-strict")); bool nvte_ck_is_v3_atomic_fp32 = getenv("NVTE_CK_IS_V3_ATOMIC_FP32", 1); int nvte_ck_how_v3_bf16_cvt = getenv("NVTE_CK_HOW_V3_BF16_CVT", 1); From 30bd46baf88b1f452f6348d48777f945c98ac6eb Mon Sep 17 00:00:00 2001 From: Ilya Panfilov Date: Tue, 27 Oct 2026 02:04:18 -0400 Subject: [PATCH 2/3] QoLA update --- 3rdparty/QoLA | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/3rdparty/QoLA b/3rdparty/QoLA index c33917c7e0..0edd909edd 160000 --- a/3rdparty/QoLA +++ b/3rdparty/QoLA @@ -1 +1 @@ -Subproject commit c33917c7e01d300615f7ef8ed355082a144bbed4 +Subproject commit 0edd909edd11e0a86b99614cfcd31f60ff845409 From c8f5ac0cc5825e4a2d5ae6aed860fa7a156f90d7 Mon Sep 17 00:00:00 2001 From: Ilya Panfilov Date: Tue, 27 Oct 2026 12:58:20 -0400 Subject: [PATCH 3/3] Update QoLA and CK-JIT --- 3rdparty/QoLA | 2 +- 3rdparty/ck_jit | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/3rdparty/QoLA b/3rdparty/QoLA index 0edd909edd..3ff1d89fd4 160000 --- a/3rdparty/QoLA +++ b/3rdparty/QoLA @@ -1 +1 @@ -Subproject commit 0edd909edd11e0a86b99614cfcd31f60ff845409 +Subproject commit 3ff1d89fd4d2204241b423de6cd3e4095202fcda diff --git a/3rdparty/ck_jit b/3rdparty/ck_jit index c0392304ae..bc28d192c4 160000 --- a/3rdparty/ck_jit +++ b/3rdparty/ck_jit @@ -1 +1 @@ -Subproject commit c0392304ae352e1cbb40f5ec0cd43ba46c5bdac2 +Subproject commit bc28d192c4f4e3a9e07fad977c22255227ad33ac