#4666·sdnext

[Issue] ROCm Unable to load file: gfx908_metadata.tn.model

Author: oOoBlackHoleSunoOoCreated Mar 1, 2026Updated Jun 23, 2026
Labelsupstreamplatformrocm

INFO:

arch: AMD64
cpu: AMD64 Family 25 Model 117 Stepping 2, AuthenticAMD
system: Windows
release: Windows-11-10.0.26200-SP0
python: 3.12.10
device: AMD Radeon 780M Graphics (1)
hip: 7.2.0


[in no particular order]

INFO     Version: app=sd.next latest=2026-02-07T12:10:52Z
INFO     ROCm: HSA_OVERRIDE_GFX_VERSION config overridden:
                         device=gfx1103 version=11.0.0
INFO     ROCm: agents=['gfx1103']
INFO     ROCm: version=7.2, using agent gfx1103
INFO     Python: version=3.12.10 platform=Windows
INFO     Torch backend: version="2.10.0+rocm7.12.0a20260228" type=ROCm HIP=7.2.0
INFO     Torch detected: gpu="AMD Radeon 780M Graphics" vram=12794 arch=(11, 0) cores=6

PROBLEM:

First of all, Text to Image works flawlessly and rocm is a massive improvement. However if i want to use impaint (also other like detailer) it seems the whole radeon driver crashes. In one instance my system even restarted. The Issue shows up on all models (Pony, SD, SDXL).

There is this repeating console message, while in the background the process seems to go on and even finishes:

MIOpen(HIP): Error [C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62] Unable to load file: gfx908_metadata.tn.model
Buffered 128 messages to file: Pchello:C:\Users\hellouser\AppData\Local\Temp\miopen_error_4592.log

So this is the inside of one of the "miopen_error_4592.log" files. I generated a few while try fixing, but i guess its per proccess run one file, so:

MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 256-512-768-3x3-128-512-768-1-1x1-1x1-1x1-0-NCHW-BF16-F in cache for file ":memory:gfx1103"
MIOpen(HIP): Error [C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62] Unable to load file: gfx908_metadata.tn.model
MIOpen(HIP): Info2 [GetSolutionsFallback] [Warning] Caught exception: (Pchello:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62: Unable to load file: gfx908_metadata.tn.model), passing empty solver vector
MIOpen(HIP): Info2 [GetSolutionsFallback] Using WTI Fallback
MIOpen(HIP): Info2 [GetSolutionsFallback] ConvDirectNaiveConvFwd Estimated WTI = 0.01
MIOpen(HIP): Info2 [GetSolutionsFallback] GemmFwdRest Estimated WTI = 0.36
MIOpen(HIP): Info2 [GetSolutionsFallback] maxSolutionCount = 1, available = 2
MIOpen(HIP): Info2 [GetSolutionsFallback] id: 85, algo: 1, time: 1000, ws: 0, name: ConvDirectNaiveConvFwd
MIOpen(HIP): Info2 [GetSolutionsFallback] id: 91, algo: 0, time: 27.7778, ws: 1811939328, name: GemmFwdRest
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 256x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault and solver GemmFwdRest
MIOpen(HIP): Info [FindSolutionImpl] GemmFwdRest (not searchable)
MIOpen(HIP): Info2 [Register] Invoker registered for algorithm 256x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault and solver GemmFwdRest
MIOpen(HIP): Info2 [SetAsFound1_0] Solver GemmFwdRest registered as find 1.0 best for miopenConvolutionFwdAlgoGEMM in 256x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault
MIOpen(HIP): Info [FindConvolution] miopenConvolutionFwdAlgoGEMM	27.7778	1811939328
MIOpen(HIP): Info [FillFindReturnParameters] FW Chosen Algorithm: GemmFwdRest , 1811939328, 27.7778
MIOpen(HIP): Info [ConvolutionForward] algo = 0, workspace = 1811939328
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 256x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault and algorithm miopenConvolutionFwdAlgoGEMM
MIOpen(HIP): Info2 [GetKernels] 0 kernels for key: miopenIm2d2Col "c256i512_768w3_3p1_1s1_1d1_1t5"
MIOpen(HIP): Info2 [AddKernel] Key: miopenIm2d2Col "c256i512_768w3_3p1_1s1_1d1_1t5 -DLOCAL_MEM_SIZE=340 -DSTRIDE_GT_1=0 -DNUM_IM_BLKS_EQ_1=0 -DUSE_IM_OFF_GUARD=1"
MIOpen(HIP): Info2 [run] kernel_name = Im2d2Col_v2, global_work_dim = { 100663296, 1, 1 }, local_work_dim = { 256, 1, 1 }
MIOpen(HIP): Info2 [CallGemm] gemm_desc: {isColMajor 0, transA 0, transB 0, m 128, n 393216, k 2304, lda 2304, ldb 393216, ldc 393216, batch_count 1, strideA 0, strideB 0, strideC 0, alpha 1, beta 0, dataType bfloat16, a_cast_type bfloat16, b_cast_type bfloat16} 
MIOpen(HIP): Info2 [CallGemm] rocBLAS
MIOpen(HIP): Info [get_device_name] Raw device name: gfx1103
MIOpen(HIP): Info [SetStream] stream: 0000000000000000, device_id: 0
MIOpen(HIP): Info [ConvolutionForward] algo = 0, workspace = 905969664
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 128x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault and algorithm miopenConvolutionFwdAlgoGEMM
MIOpen(HIP): Info2 [GetKernels] 0 kernels for key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5"
MIOpen(HIP): Info2 [AddKernel] Key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5 -DLOCAL_MEM_SIZE=340 -DSTRIDE_GT_1=0 -DNUM_IM_BLKS_EQ_1=0 -DUSE_IM_OFF_GUARD=1"
MIOpen(HIP): Info2 [run] kernel_name = Im2d2Col_v2, global_work_dim = { 50331648, 1, 1 }, local_work_dim = { 256, 1, 1 }
MIOpen(HIP): Info2 [CallGemm] gemm_desc: {isColMajor 0, transA 0, transB 0, m 128, n 393216, k 1152, lda 1152, ldb 393216, ldc 393216, batch_count 1, strideA 0, strideB 0, strideC 0, alpha 1, beta 0, dataType bfloat16, a_cast_type bfloat16, b_cast_type bfloat16} 
MIOpen(HIP): Info2 [CallGemm] rocBLAS
MIOpen(HIP): Info [get_device_name] Raw device name: gfx1103
MIOpen(HIP): Info [SetStream] stream: 0000000000000000, device_id: 0
MIOpen(HIP): Info2 [GetWorkSpaceSize] 
MIOpen(HIP): Info [GetSolutions] 
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 256-512-768-1x1-128-512-768-1-0x0-1x1-1x1-0-NCHW-BF16-F in cache for file "C:\\Users\\hellouser\\.miopen\\db\\gfx1103_6.HIP.3_5_1_9e9e9009cd.ufdb.txt"
MIOpen(HIP): Info2 [FindRecord] Looking for key 256-512-768-1x1-128-512-768-1-0x0-1x1-1x1-0-NCHW-BF16-F in file ""
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 256-512-768-1x1-128-512-768-1-0x0-1x1-1x1-0-NCHW-BF16-F in cache for file ":memory:gfx1103"
MIOpen(HIP): Error [C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62] Unable to load file: gfx908_metadata.tn.model
MIOpen(HIP): Info2 [GetSolutionsFallback] [Warning] Caught exception: (Pchello:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62: Unable to load file: gfx908_metadata.tn.model), passing empty solver vector
MIOpen(HIP): Info2 [GetSolutionsFallback] Using WTI Fallback
MIOpen(HIP): Info2 [GetSolutionsFallback] ConvDirectNaiveConvFwd Estimated WTI = 0.01
MIOpen(HIP): Info2 [GetSolutionsFallback] GemmFwd1x1_0_1 Estimated WTI = 0.9
MIOpen(HIP): Info2 [GetSolutionsFallback] maxSolutionCount = 1, available = 2
MIOpen(HIP): Info2 [GetSolutionsFallback] id: 85, algo: 1, time: 1000, ws: 0, name: ConvDirectNaiveConvFwd
MIOpen(HIP): Info2 [GetSolutionsFallback] id: 88, algo: 0, time: 11.1111, ws: 0, name: GemmFwd1x1_0_1
MIOpen(HIP): Info [GetWorkSpaceSize] 0
MIOpen(HIP): Info [FindConvFwdAlgorithm] requestAlgoCount = 1, workspace = 0
MIOpen(HIP): Info [GetSolutions] 
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 256-512-768-1x1-128-512-768-1-0x0-1x1-1x1-0-NCHW-BF16-F in cache for file "C:\\Users\\hellouser\\.miopen\\db\\gfx1103_6.HIP.3_5_1_9e9e9009cd.ufdb.txt"
MIOpen(HIP): Info2 [FindRecord] Looking for key 256-512-768-1x1-128-512-768-1-0x0-1x1-1x1-0-NCHW-BF16-F in file ""
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 256-512-768-1x1-128-512-768-1-0x0-1x1-1x1-0-NCHW-BF16-F in cache for file ":memory:gfx1103"
MIOpen(HIP): Error [C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62] Unable to load file: gfx908_metadata.tn.model
MIOpen(HIP): Info2 [GetSolutionsFallback] [Warning] Caught exception: (Pchello:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62: Unable to load file: gfx908_metadata.tn.model), passing empty solver vector
MIOpen(HIP): Info2 [GetSolutionsFallback] Using WTI Fallback
MIOpen(HIP): Info2 [GetSolutionsFallback] ConvDirectNaiveConvFwd Estimated WTI = 0.01
MIOpen(HIP): Info2 [GetSolutionsFallback] GemmFwd1x1_0_1 Estimated WTI = 0.9
MIOpen(HIP): Info2 [GetSolutionsFallback] maxSolutionCount = 1, available = 2
MIOpen(HIP): Info2 [GetSolutionsFallback] id: 85, algo: 1, time: 1000, ws: 0, name: ConvDirectNaiveConvFwd
MIOpen(HIP): Info2 [GetSolutionsFallback] id: 88, algo: 0, time: 11.1111, ws: 0, name: GemmFwd1x1_0_1
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 256x512x768x1x1x128x512x768x1xNCHWxBF16x0x0x1x1x1x1x1xFxDefault and solver GemmFwd1x1_0_1
MIOpen(HIP): Info [FindSolutionImpl] GemmFwd1x1_0_1 (not searchable)
MIOpen(HIP): Info2 [Register] Invoker registered for algorithm 256x512x768x1x1x128x512x768x1xNCHWxBF16x0x0x1x1x1x1x1xFxDefault and solver GemmFwd1x1_0_1
MIOpen(HIP): Info2 [SetAsFound1_0] Solver GemmFwd1x1_0_1 registered as find 1.0 best for miopenConvolutionFwdAlgoGEMM in 256x512x768x1x1x128x512x768x1xNCHWxBF16x0x0x1x1x1x1x1xFxDefault
MIOpen(HIP): Info [FindConvolution] miopenConvolutionFwdAlgoGEMM	11.1111	0
MIOpen(HIP): Info [FillFindReturnParameters] FW Chosen Algorithm: GemmFwd1x1_0_1 , 0, 11.1111
MIOpen(HIP): Info [ConvolutionForward] algo = 0, workspace = 0
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 256x512x768x1x1x128x512x768x1xNCHWxBF16x0x0x1x1x1x1x1xFxDefault and algorithm miopenConvolutionFwdAlgoGEMM
MIOpen(HIP): Info2 [CallGemmStridedBatched] gemm_desc: {isColMajor 0, transA 0, transB 0, m 128, n 393216, k 256, lda 256, ldb 393216, ldc 393216, batch_count 1, strideA 0, strideB 100663296, strideC 50331648, alpha 1, beta 0, dataType bfloat16, a_cast_type bfloat16, b_cast_type bfloat16} 
MIOpen(HIP): Info2 [CallGemmStridedBatched] rocBLAS
MIOpen(HIP): Info [get_device_name] Raw device name: gfx1103
MIOpen(HIP): Info [SetStream] stream: 0000000000000000, device_id: 0
MIOpen(HIP): Info [ConvolutionForward] algo = 0, workspace = 905969664
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 128x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault and algorithm miopenConvolutionFwdAlgoGEMM
MIOpen(HIP): Info2 [GetKernels] 0 kernels for key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5"
MIOpen(HIP): Info2 [AddKernel] Key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5 -DLOCAL_MEM_SIZE=340 -DSTRIDE_GT_1=0 -DNUM_IM_BLKS_EQ_1=0 -DUSE_IM_OFF_GUARD=1"
MIOpen(HIP): Info2 [run] kernel_name = Im2d2Col_v2, global_work_dim = { 50331648, 1, 1 }, local_work_dim = { 256, 1, 1 }
MIOpen(HIP): Info2 [CallGemm] gemm_desc: {isColMajor 0, transA 0, transB 0, m 128, n 393216, k 1152, lda 1152, ldb 393216, ldc 393216, batch_count 1, strideA 0, strideB 0, strideC 0, alpha 1, beta 0, dataType bfloat16, a_cast_type bfloat16, b_cast_type bfloat16} 
MIOpen(HIP): Info2 [CallGemm] rocBLAS
MIOpen(HIP): Info [get_device_name] Raw device name: gfx1103
MIOpen(HIP): Info [SetStream] stream: 0000000000000000, device_id: 0
MIOpen(HIP): Info [ConvolutionForward] algo = 0, workspace = 905969664
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 128x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault and algorithm miopenConvolutionFwdAlgoGEMM
MIOpen(HIP): Info2 [GetKernels] 0 kernels for key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5"
MIOpen(HIP): Info2 [AddKernel] Key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5 -DLOCAL_MEM_SIZE=340 -DSTRIDE_GT_1=0 -DNUM_IM_BLKS_EQ_1=0 -DUSE_IM_OFF_GUARD=1"
MIOpen(HIP): Info2 [run] kernel_name = Im2d2Col_v2, global_work_dim = { 50331648, 1, 1 }, local_work_dim = { 256, 1, 1 }
MIOpen(HIP): Info2 [CallGemm] gemm_desc: {isColMajor 0, transA 0, transB 0, m 128, n 393216, k 1152, lda 1152, ldb 393216, ldc 393216, batch_count 1, strideA 0, strideB 0, strideC 0, alpha 1, beta 0, dataType bfloat16, a_cast_type bfloat16, b_cast_type bfloat16} 
MIOpen(HIP): Info2 [CallGemm] rocBLAS
MIOpen(HIP): Info [get_device_name] Raw device name: gfx1103
MIOpen(HIP): Info [SetStream] stream: 0000000000000000, device_id: 0
MIOpen(HIP): Info [ConvolutionForward] algo = 0, workspace = 905969664
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 128x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault and algorithm miopenConvolutionFwdAlgoGEMM
MIOpen(HIP): Info2 [GetKernels] 0 kernels for key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5"
MIOpen(HIP): Info2 [AddKernel] Key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5 -DLOCAL_MEM_SIZE=340 -DSTRIDE_GT_1=0 -DNUM_IM_BLKS_EQ_1=0 -DUSE_IM_OFF_GUARD=1"
MIOpen(HIP): Info2 [run] kernel_name = Im2d2Col_v2, global_work_dim = { 50331648, 1, 1 }, local_work_dim = { 256, 1, 1 }
MIOpen(HIP): Info2 [CallGemm] gemm_desc: {isColMajor 0, transA 0, transB 0, m 128, n 393216, k 1152, lda 1152, ldb 393216, ldc 393216, batch_count 1, strideA 0, strideB 0, strideC 0, alpha 1, beta 0, dataType bfloat16, a_cast_type bfloat16, b_cast_type bfloat16} 
MIOpen(HIP): Info2 [CallGemm] rocBLAS
MIOpen(HIP): Info [get_device_name] Raw device name: gfx1103
MIOpen(HIP): Info [SetStream] stream: 0000000000000000, device_id: 0
MIOpen(HIP): Info [ConvolutionForward] algo = 0, workspace = 905969664
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 128x512x768x3x3x128x512x768x1xNCHWxBF16x1x1x1x1x1x1x1xFxDefault and algorithm miopenConvolutionFwdAlgoGEMM
MIOpen(HIP): Info2 [GetKernels] 0 kernels for key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5"
MIOpen(HIP): Info2 [AddKernel] Key: miopenIm2d2Col "c128i512_768w3_3p1_1s1_1d1_1t5 -DLOCAL_MEM_SIZE=340 -DSTRIDE_GT_1=0 -DNUM_IM_BLKS_EQ_1=0 -DUSE_IM_OFF_GUARD=1"
MIOpen(HIP): Info2 [run] kernel_name = Im2d2Col_v2, global_work_dim = { 50331648, 1, 1 }, local_work_dim = { 256, 1, 1 }
MIOpen(HIP): Info2 [CallGemm] gemm_desc: {isColMajor 0, transA 0, transB 0, m 128, n 393216, k 1152, lda 1152, ldb 393216, ldc 393216, batch_count 1, strideA 0, strideB 0, strideC 0, alpha 1, beta 0, dataType bfloat16, a_cast_type bfloat16, b_cast_type bfloat16} 
MIOpen(HIP): Info2 [CallGemm] rocBLAS
MIOpen(HIP): Info [get_device_name] Raw device name: gfx1103
MIOpen(HIP): Info [SetStream] stream: 0000000000000000, device_id: 0
MIOpen(HIP): Info2 [GetWorkSpaceSize] 
MIOpen(HIP): Info [GetSolutions] 
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 128-512-768-3x3-3-512-768-1-1x1-1x1-1x1-0-NCHW-BF16-F in cache for file "C:\\Users\\hellouser\\.miopen\\db\\gfx1103_6.HIP.3_5_1_9e9e9009cd.ufdb.txt"
MIOpen(HIP): Info2 [FindRecord] Looking for key 128-512-768-3x3-3-512-768-1-1x1-1x1-1x1-0-NCHW-BF16-F in file ""
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 128-512-768-3x3-3-512-768-1-1x1-1x1-1x1-0-NCHW-BF16-F in cache for file ":memory:gfx1103"
MIOpen(HIP): Error [C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62] Unable to load file: gfx908_metadata.tn.model
MIOpen(HIP): Info2 [GetSolutionsFallback] [Warning] Caught exception: (Pchello:C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62: Unable to load file: gfx908_metadata.tn.model), passing empty solver vector
MIOpen(HIP): Info2 [GetSolutionsFallback] Using WTI Fallback
MIOpen(HIP): Info2 [GetSolutionsFallback] ConvDirectNaiveConvFwd Estimated WTI = 0.01
MIOpen(HIP): Info2 [GetSolutionsFallback] GemmFwdRest Estimated WTI = 0.36
MIOpen(HIP): Info2 [GetSolutionsFallback] maxSolutionCount = 1, available = 2
MIOpen(HIP): Info2 [GetSolutionsFallback] id: 85, algo: 1, time: 1000, ws: 0, name: ConvDirectNaiveConvFwd
MIOpen(HIP): Info2 [GetSolutionsFallback] id: 91, algo: 0, time: 27.7778, ws: 905969664, name: GemmFwdRest
MIOpen(HIP): Info [GetWorkSpaceSize] 905969664
MIOpen(HIP): Info [FindConvFwdAlgorithm] requestAlgoCount = 1, workspace = 905969664
MIOpen(HIP): Info [GetSolutions] 
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 128-512-768-3x3-3-512-768-1-1x1-1x1-1x1-0-NCHW-BF16-F in cache for file "C:\\Users\\hellouser\\.miopen\\db\\gfx1103_6.HIP.3_5_1_9e9e9009cd.ufdb.txt"
MIOpen(HIP): Info2 [FindRecord] Looking for key 128-512-768-3x3-3-512-768-1-1x1-1x1-1x1-0-NCHW-BF16-F in file ""
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 128-512-768-3x3-3-512-768-1-1x1-1x1-1x1-0-NCHW-BF16-F in cache for file ":memory:gfx1103"
MIOpen(HIP): Error [C:/home/runner/_work/TheRock/TheRock/rocm-libraries/projects/miopen/src/include\miopen/conv/heuristics/ai_heuristics.hpp:62] Unable to load file: gfx908_metadata.tn.model

What i tried to fix:

  1. starting sdnext in lowvram, medvram mode
  2. There were some post online with kind a the same issue but for other amd cards so i tried to adjust my webui-user.bat:
  • They recommend HIP_VISIBLE_DEVICES
  • They recommend to set TORCH_USE_HIP_DSA & AMD_SERIALIZE_KERNEL for debugging due to some async issues in the logs? The log is may not in order otherwise.
  • The gfx908_metadata.tn.model file is in my venv\Lib\site-packages\_rocm_sdk_devel\bin folder somebody mentioned to set HIPBLASLT_TENSILE_LIBPATH
set COMMANDLINE_ARGS= --use-rocm --medvram --debug --locale en
set TORCH_USE_HIP_DSA=1
set AMD_SERIALIZE_KERNEL=3
set HIP_VISIBLE_DEVICES=0
set HIPBLASLT_TENSILE_LIBPATH=D:\Atlas\repos\sdnext_rocm2\sdnext\venv\Lib\site-packages\_rocm_sdk_devel\bin

Unfortunately, neither worked; I'm still getting the same error message.