Issue Station
Sign In
Tools hereToolbox

pytorch - 💡(How to fix) Fix hipBLASLt path ignores backward fp16_alt_impl for Half GEMMs, causing FTZ in backward and silent training divergence on gfx90a

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…

Error Message

  • PyTorch 2.3.0 + ROCm 6.0: default backend is rocBLAS, backward result matches ref.
  • PyTorch 2.9.1 + ROCm 6.4: default backend is hipBLASLt, backward result is 0.0.
  • PyTorch 2.9.1 + ROCm 6.4 with forced rocBLAS: backward result again matches ref.

Root Cause

Likely root cause

Fix Action

Fix / Workaround

With BLAS logging enabled, rocBLAS backward GEMMs are dispatched with flags 4 (fp16_alt_impl), while hipBLASLt forward/backward matmuls look identical apart from transpose, suggesting the backward guard is not wired through the hipBLASLt path.

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 48 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 128 On-line CPU(s) list: 0-127 Vendor ID: AuthenticAMD Model name: AMD EPYC 7A53 64-Core Processor CPU family: 25 Model: 48 Thread(s) per core: 2 Core(s) per socket: 64 Socket(s): 1 Stepping: 1 Frequency boost: enabled CPU max MHz: 3541.0149 CPU min MHz: 1500.0000 BogoMIPS: 3992.69 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm ibpb_exit_to_user Virtualization: AMD-V L1d cache: 2 MiB (64 instances) L1i cache: 2 MiB (64 instances) L2 cache: 32 MiB (64 instances) L3 cache: 256 MiB (8 instances) NUMA node(s): 4 NUMA node0 CPU(s): 0-15,64-79 NUMA node1 CPU(s): 16-31,80-95 NUMA node2 CPU(s): 32-47,96-111 NUMA node3 CPU(s): 48-63,112-127 Vulnerability Gather data sampling: Not affected Vulnerability Indirect target selection: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Mitigation; Safe RET Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsa: Vulnerable: No microcode Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB before exit to userspace

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 48 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 128 On-line CPU(s) list: 0-127 Vendor ID: AuthenticAMD Model name: AMD EPYC 7A53 64-Core Processor CPU family: 25 Model: 48 Thread(s) per core: 2 Core(s) per socket: 64 Socket(s): 1 Stepping: 1 Frequency boost: enabled CPU(s) scaling MHz: 56% CPU max MHz: 3541.0149 CPU min MHz: 1500.0000 BogoMIPS: 3992.69 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm ibpb_exit_to_user Virtualization: AMD-V L1d cache: 2 MiB (64 instances) L1i cache: 2 MiB (64 instances) L2 cache: 32 MiB (64 instances) L3 cache: 256 MiB (8 instances) NUMA node(s): 4 NUMA node0 CPU(s): 0-15,64-79 NUMA node1 CPU(s): 16-31,80-95 NUMA node2 CPU(s): 32-47,96-111 NUMA node3 CPU(s): 48-63,112-127 Vulnerability Gather data sampling: Not affected Vulnerability Indirect target selection: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Mitigation; Safe RET Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsa: Vulnerable: No microcode Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB before exit to userspace

Code Example

torch.backends.cuda.preferred_blas_library("cublas")

---

import torch

dtype = torch.float16
M = K = N = 64
sub_val = torch.finfo(dtype).tiny / 2
ref = M * sub_val

def check_backward(force_rocblas=False):
    if force_rocblas:
        torch.backends.cuda.preferred_blas_library("cublas")

    print("PyTorch:", torch.__version__)
    print("ROCm:", torch.version.hip)
    if hasattr(torch.backends.cuda, "preferred_blas_library"):
        print("preferred_blas_library:", torch.backends.cuda.preferred_blas_library())

    x = torch.ones(M, K, dtype=dtype, device="cuda")
    w = torch.nn.Parameter(torch.ones(K, N, dtype=dtype, device="cuda"))
    d = torch.full((M, N), sub_val, dtype=dtype, device="cuda")

    (x @ w).backward(d)
    torch.cuda.synchronize()
    val = w.grad[0, 0].item()
    print("grad[0,0] =", val, "expected =", ref)

print("Default backend")
check_backward(force_rocblas=False)

print("\nForced rocBLAS")
check_backward(force_rocblas=True)

---

PyTorch version: 2.3.0+rocm6.0
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.0.32830-d62f6a171

OS: SUSE Linux Enterprise Server 15 SP5 (x86_64)
GCC version: Could not collect
Clang version: 17.0.0 (https://github.com/RadeonOpenCompute/llvm-project roc-6.0.3 24012 af27734ed982b52a9f1be0f035ac91726fc697e4)
CMake version: version 4.1.2
Libc version: glibc-2.31

Python version: 3.12.12 | packaged by conda-forge | (main, Oct 13 2025, 14:34:15) [GCC 14.3.0] (64-bit runtime)
Python platform: Linux-6.4.0-150600.23.73_15.0.14-cray_shasta_c-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: AMD Instinct MI250X (gfx90a:sramecc+:xnack-)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: 6.0.32830
MIOpen runtime version: 3.0.0
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  128
On-line CPU(s) list:                     0-127
Vendor ID:                               AuthenticAMD
Model name:                              AMD EPYC 7A53 64-Core Processor
CPU family:                              25
Model:                                   48
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               1
Stepping:                                1
Frequency boost:                         enabled
CPU max MHz:                             3541.0149
CPU min MHz:                             1500.0000
BogoMIPS:                                3992.69
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm ibpb_exit_to_user
Virtualization:                          AMD-V
L1d cache:                               2 MiB (64 instances)
L1i cache:                               2 MiB (64 instances)
L2 cache:                                32 MiB (64 instances)
L3 cache:                                256 MiB (8 instances)
NUMA node(s):                            4
NUMA node0 CPU(s):                       0-15,64-79
NUMA node1 CPU(s):                       16-31,80-95
NUMA node2 CPU(s):                       32-47,96-111
NUMA node3 CPU(s):                       48-63,112-127
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; Safe RET
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Vulnerable: No microcode
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] pytorch-lightning==2.5.5
[pip3] pytorch-triton-rocm==3.0.0
[pip3] torch==2.3.0+rocm6.0
[pip3] torch-geometric==2.6.1
[pip3] torchaudio==2.3.0+rocm6.0
[pip3] torchinfo==1.8.0
[pip3] torchmetrics==1.8.2
[pip3] torchvision==0.18.0+rocm6.0
[conda] numpy                                 1.26.4               pypi_0              pypi
[conda] pytorch-lightning                     2.5.5                pypi_0              pypi
[conda] pytorch-triton-rocm                   3.0.0                pypi_0              pypi
[conda] torch                                 2.3.0+rocm6.0        pypi_0              pypi
[conda] torch-geometric                       2.6.1                pypi_0              pypi
[conda] torchaudio                            2.3.0+rocm6.0        pypi_0              pypi
[conda] torchinfo                             1.8.0                pypi_0              pypi
[conda] torchmetrics                          1.8.2                pypi_0              pypi
[conda] torchvision                           0.18.0+rocm6.0       pypi_0

---

PyTorch version: 2.9.1+rocm6.4
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.4.43484-123eb5128

OS: Ubuntu 24.04.4 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.39

Python version: 3.12.3 (main, Jan 22 2026, 20:57:42) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.4.0-150600.23.73_15.0.14-cray_shasta_c-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: 
GPU models and configuration: AMD Instinct MI250X (gfx90a:sramecc+:xnack-)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: 6.4.43484
MIOpen runtime version: 3.4.0
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  128
On-line CPU(s) list:                     0-127
Vendor ID:                               AuthenticAMD
Model name:                              AMD EPYC 7A53 64-Core Processor
CPU family:                              25
Model:                                   48
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               1
Stepping:                                1
Frequency boost:                         enabled
CPU(s) scaling MHz:                      56%
CPU max MHz:                             3541.0149
CPU min MHz:                             1500.0000
BogoMIPS:                                3992.69
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm ibpb_exit_to_user
Virtualization:                          AMD-V
L1d cache:                               2 MiB (64 instances)
L1i cache:                               2 MiB (64 instances)
L2 cache:                                32 MiB (64 instances)
L3 cache:                                256 MiB (8 instances)
NUMA node(s):                            4
NUMA node0 CPU(s):                       0-15,64-79
NUMA node1 CPU(s):                       16-31,80-95
NUMA node2 CPU(s):                       32-47,96-111
NUMA node3 CPU(s):                       48-63,112-127
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; Safe RET
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Vulnerable: No microcode
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

Versions of relevant libraries:
[pip3] numpy==2.4.2
[pip3] pytorch-triton-rocm==3.5.1
[pip3] torch==2.9.1+rocm6.4
[pip3] torchaudio==2.9.1+rocm6.4
[pip3] torchvision==0.24.1+rocm6.4
[conda] Could not collect
RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

Summary

On AMD gfx90a (MI250X), PyTorch 2.7+ defaults blasPreferredBackend() to Cublaslt on ROCm. On this path, backward float16 GEMMs appear to run without the ROCm backward-pass alternative implementation that preserves fp16 subnormals. As a result, subnormal fp16 inputs in backward GEMMs are flushed to zero, which can silently corrupt gradients and cause training divergence.

Forcing rocBLAS with:

torch.backends.cuda.preferred_blas_library("cublas")

restores the expected behavior on the same system.

Minimal reproducer

import torch

dtype = torch.float16
M = K = N = 64
sub_val = torch.finfo(dtype).tiny / 2
ref = M * sub_val

def check_backward(force_rocblas=False):
    if force_rocblas:
        torch.backends.cuda.preferred_blas_library("cublas")

    print("PyTorch:", torch.__version__)
    print("ROCm:", torch.version.hip)
    if hasattr(torch.backends.cuda, "preferred_blas_library"):
        print("preferred_blas_library:", torch.backends.cuda.preferred_blas_library())

    x = torch.ones(M, K, dtype=dtype, device="cuda")
    w = torch.nn.Parameter(torch.ones(K, N, dtype=dtype, device="cuda"))
    d = torch.full((M, N), sub_val, dtype=dtype, device="cuda")

    (x @ w).backward(d)
    torch.cuda.synchronize()
    val = w.grad[0, 0].item()
    print("grad[0,0] =", val, "expected =", ref)

print("Default backend")
check_backward(force_rocblas=False)

print("\nForced rocBLAS")
check_backward(force_rocblas=True)

Observed behavior

  • PyTorch 2.3.0 + ROCm 6.0: default backend is rocBLAS, backward result matches ref.
  • PyTorch 2.9.1 + ROCm 6.4: default backend is hipBLASLt, backward result is 0.0.
  • PyTorch 2.9.1 + ROCm 6.4 with forced rocBLAS: backward result again matches ref.

With BLAS logging enabled, rocBLAS backward GEMMs are dispatched with flags 4 (fp16_alt_impl), while hipBLASLt forward/backward matmuls look identical apart from transpose, suggesting the backward guard is not wired through the hipBLASLt path.

Expected behavior

One of these should hold:

  1. Backward fp16 GEMMs on ROCm should preserve the current rocBLAS behavior when PyTorch selects hipBLASLt by default.
  2. If hipBLASLt cannot support the backward-safe path yet, gfx90a should not default to Cublaslt for fp16 training workloads where this changes numerical behavior.

Likely root cause

ROCmBackwardPassGuard is honored on the rocBLAS path, where PyTorch passes rocblas_gemm_flags_fp16_alt_impl, but not on gemm_internal_cublaslt<at::Half>(). The PyTorch numerical accuracy notes state that PyTorch uses alternate implementations during backward for FP16 denorm handling, but that no longer seems true on gfx90a once the default backend switched to hipBLASLt in 2.7.0.

Versions

Pytorch 2.3.0 + ROCm6.0

PyTorch version: 2.3.0+rocm6.0
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.0.32830-d62f6a171

OS: SUSE Linux Enterprise Server 15 SP5 (x86_64)
GCC version: Could not collect
Clang version: 17.0.0 (https://github.com/RadeonOpenCompute/llvm-project roc-6.0.3 24012 af27734ed982b52a9f1be0f035ac91726fc697e4)
CMake version: version 4.1.2
Libc version: glibc-2.31

Python version: 3.12.12 | packaged by conda-forge | (main, Oct 13 2025, 14:34:15) [GCC 14.3.0] (64-bit runtime)
Python platform: Linux-6.4.0-150600.23.73_15.0.14-cray_shasta_c-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: AMD Instinct MI250X (gfx90a:sramecc+:xnack-)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: 6.0.32830
MIOpen runtime version: 3.0.0
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  128
On-line CPU(s) list:                     0-127
Vendor ID:                               AuthenticAMD
Model name:                              AMD EPYC 7A53 64-Core Processor
CPU family:                              25
Model:                                   48
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               1
Stepping:                                1
Frequency boost:                         enabled
CPU max MHz:                             3541.0149
CPU min MHz:                             1500.0000
BogoMIPS:                                3992.69
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm ibpb_exit_to_user
Virtualization:                          AMD-V
L1d cache:                               2 MiB (64 instances)
L1i cache:                               2 MiB (64 instances)
L2 cache:                                32 MiB (64 instances)
L3 cache:                                256 MiB (8 instances)
NUMA node(s):                            4
NUMA node0 CPU(s):                       0-15,64-79
NUMA node1 CPU(s):                       16-31,80-95
NUMA node2 CPU(s):                       32-47,96-111
NUMA node3 CPU(s):                       48-63,112-127
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; Safe RET
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Vulnerable: No microcode
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] pytorch-lightning==2.5.5
[pip3] pytorch-triton-rocm==3.0.0
[pip3] torch==2.3.0+rocm6.0
[pip3] torch-geometric==2.6.1
[pip3] torchaudio==2.3.0+rocm6.0
[pip3] torchinfo==1.8.0
[pip3] torchmetrics==1.8.2
[pip3] torchvision==0.18.0+rocm6.0
[conda] numpy                                 1.26.4               pypi_0              pypi
[conda] pytorch-lightning                     2.5.5                pypi_0              pypi
[conda] pytorch-triton-rocm                   3.0.0                pypi_0              pypi
[conda] torch                                 2.3.0+rocm6.0        pypi_0              pypi
[conda] torch-geometric                       2.6.1                pypi_0              pypi
[conda] torchaudio                            2.3.0+rocm6.0        pypi_0              pypi
[conda] torchinfo                             1.8.0                pypi_0              pypi
[conda] torchmetrics                          1.8.2                pypi_0              pypi
[conda] torchvision                           0.18.0+rocm6.0       pypi_0

Pytorch2.9.1+Rocm6.4

PyTorch version: 2.9.1+rocm6.4
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.4.43484-123eb5128

OS: Ubuntu 24.04.4 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.39

Python version: 3.12.3 (main, Jan 22 2026, 20:57:42) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.4.0-150600.23.73_15.0.14-cray_shasta_c-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: 
GPU models and configuration: AMD Instinct MI250X (gfx90a:sramecc+:xnack-)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: 6.4.43484
MIOpen runtime version: 3.4.0
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  128
On-line CPU(s) list:                     0-127
Vendor ID:                               AuthenticAMD
Model name:                              AMD EPYC 7A53 64-Core Processor
CPU family:                              25
Model:                                   48
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               1
Stepping:                                1
Frequency boost:                         enabled
CPU(s) scaling MHz:                      56%
CPU max MHz:                             3541.0149
CPU min MHz:                             1500.0000
BogoMIPS:                                3992.69
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm ibpb_exit_to_user
Virtualization:                          AMD-V
L1d cache:                               2 MiB (64 instances)
L1i cache:                               2 MiB (64 instances)
L2 cache:                                32 MiB (64 instances)
L3 cache:                                256 MiB (8 instances)
NUMA node(s):                            4
NUMA node0 CPU(s):                       0-15,64-79
NUMA node1 CPU(s):                       16-31,80-95
NUMA node2 CPU(s):                       32-47,96-111
NUMA node3 CPU(s):                       48-63,112-127
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; Safe RET
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Vulnerable: No microcode
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

Versions of relevant libraries:
[pip3] numpy==2.4.2
[pip3] pytorch-triton-rocm==3.5.1
[pip3] torch==2.9.1+rocm6.4
[pip3] torchaudio==2.9.1+rocm6.4
[pip3] torchvision==0.24.1+rocm6.4
[conda] Could not collect

cc @ezyang @albanD @gqchen @nikitaved @soulitzer @Varal7 @bobrenjc93 @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @jianyuh @mruberry @walterddr @xwang233 @Lezcano

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

FAQ

Expected behavior

One of these should hold:

  1. Backward fp16 GEMMs on ROCm should preserve the current rocBLAS behavior when PyTorch selects hipBLASLt by default.
  2. If hipBLASLt cannot support the backward-safe path yet, gfx90a should not default to Cublaslt for fp16 training workloads where this changes numerical behavior.
#api#vector store#embedding generation#cache error#pipeline error

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

pytorch - 💡(How to fix) Fix hipBLASLt path ignores backward fp16_alt_impl for Half GEMMs, causing FTZ in backward and silent training divergence on gfx90a