Benchmark · pre-generated, always available← back to leaderboard

Migration Readiness Report

state-spaces/mamba

88/ 100 ready

Manual blockers present

state-spaces/mamba scored 88/100 for ROCm readiness (Manual blockers present). Of 268 CUDA-related findings, 203 work as-is on ROCm PyTorch, 63 need a mechanical HIPify-style change, and 2 are manual blockers. Focus first on the manual blockers — they gate the migration.

203Works as-is

63Mechanical change

2Manual blocker

Findings

268

Python files

Files scanned

107

Custom CUDA kernels

detected

Detected but out of scope (not analyzed): C++, C/C++ header

Findings by file

268 findings · 57 files

mamba_ssm/ops/cute/mamba3/mamba3_step_fn.py· 3

cuda-python low-level driver API

import cuda.bindings.driver as cuda

Fireworks AI:12

cuda-python (from cuda import cudart/cuda) wraps the CUDA driver directly. There is no drop-in ROCm binding; the driver calls must be rewritten against HIP.

RecommendRewrite driver-API calls against the HIP runtime.

Suggested change · advisory

# mamba_ssm/ops/cute/mamba3/mamba3_step_fn.py:12
- import cuda.bindings.driver as cuda
+ import cuda.bindings.driver as cuda  # advisory: see ROCm note above

Device string "cuda"

device = torch.device("cuda:0")

template:823

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Triton dependency

from triton.testing import do_bench_cudagraph

Fireworks AI:930

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Suggested change · advisory

# mamba_ssm/ops/cute/mamba3/mamba3_step_fn.py:930
- from triton.testing import do_bench_cudagraph
+ from triton.testing import do_bench_cudagraph  # advisory: see ROCm note above

setup.py· 1

Inline PTX assembly

"--ptxas-options=-v",

Fireworks AI:240

Inline PTX is NVIDIA ISA and cannot run on AMD. The block must be hand-rewritten in HIP/GCN or replaced with a portable path.

RecommendRewrite the PTX block in HIP or a portable high-level equivalent.

Suggested change · advisory

# setup.py:240
- "--ptxas-options=-v",
+ "--ptxas-options=-v",  # advisory: see ROCm note above

csrc/selective_scan/reverse_scan.cuh· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/reverse_scan.cuh:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_bwd_bf16_complex.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_bwd_bf16_complex.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_bwd_bf16_real.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_bwd_bf16_real.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_bwd_fp16_complex.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_bwd_fp16_complex.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_bwd_fp16_real.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_bwd_fp16_real.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_bwd_fp32_complex.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_bwd_fp32_complex.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_bwd_fp32_real.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_bwd_fp32_real.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_bwd_kernel.cuh· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_bwd_kernel.cuh:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_fwd_bf16.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_fwd_bf16.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_fwd_fp16.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_fwd_fp16.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_fwd_fp32.cu· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_fwd_fp32.cu:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/selective_scan_fwd_kernel.cuh· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/selective_scan_fwd_kernel.cuh:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

csrc/selective_scan/uninitialized_copy.cuh· 1

Custom CUDA kernels (.cu/.cuh)

CUDA kernel source

Fireworks AIfile

Hand-written .cu/.cuh kernels compile only with nvcc. AMD's HIPify tools translate the vast majority of CUDA C++ to HIP automatically.

RecommendRun hipify-perl over the .cu/.cuh sources and build with hipcc.

Suggested change · advisory

# csrc/selective_scan/uninitialized_copy.cuh:?
- CUDA kernel source
+ CUDA kernel source  # advisory: see ROCm note above

mamba_ssm/modules/mha.py· 2

FlashAttention dependency

from flash_attn import flash_attn_with_kvcache

Fireworks AI:11

FlashAttention has an official ROCm/CK implementation. Swapping the wheel (or using PyTorch SDPA) is mechanical.

RecommendInstall the ROCm FlashAttention build or fall back to torch SDPA.

Suggested change · advisory

# mamba_ssm/modules/mha.py:11
- from flash_attn import flash_attn_with_kvcache
+ from flash_attn import flash_attn_with_kvcache  # advisory: see ROCm note above

FlashAttention dependency

from flash_attn.layers.rotary import RotaryEmbedding

Fireworks AI:16

FlashAttention has an official ROCm/CK implementation. Swapping the wheel (or using PyTorch SDPA) is mechanical.

RecommendInstall the ROCm FlashAttention build or fall back to torch SDPA.

Suggested change · advisory

# mamba_ssm/modules/mha.py:16
- from flash_attn.layers.rotary import RotaryEmbedding
+ from flash_attn.layers.rotary import RotaryEmbedding  # advisory: see ROCm note above

mamba_ssm/ops/tilelang/mamba3/mamba3_mimo_bwd_varlen.py· 1

Triton dependency

from triton.testing import do_bench

Fireworks AI:37

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Suggested change · advisory

# mamba_ssm/ops/tilelang/mamba3/mamba3_mimo_bwd_varlen.py:37
- from triton.testing import do_bench
+ from triton.testing import do_bench  # advisory: see ROCm note above

mamba_ssm/ops/tilelang/mamba3/mamba3_mimo_bwd.py· 1

Triton dependency

from triton.testing import do_bench

Fireworks AI:13

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Suggested change · advisory

# mamba_ssm/ops/tilelang/mamba3/mamba3_mimo_bwd.py:13
- from triton.testing import do_bench
+ from triton.testing import do_bench  # advisory: see ROCm note above

mamba_ssm/ops/triton/k_activations.py· 4

Triton dependency

import triton

template:5

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:6

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:60

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:146

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/ops/triton/layer_norm.py· 7

Triton dependency

import triton

template:16

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:17

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

gcn_arch_name = torch.cuda.get_device_properties(0).gcnArchName

template:135

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

device_name = torch.cuda.get_device_properties(0).name

template:144

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:370

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

sm_count = torch.cuda.get_device_properties(x.device).multi_processor_count

template:660

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:671

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/ops/triton/layernorm_gated.py· 5

Triton dependency

import triton

template:12

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:13

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:139

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

sm_count = torch.cuda.get_device_properties(x.device).multi_processor_count

template:309

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:317

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/ops/triton/mamba3/angle_dt.py· 2

Triton dependency

import triton

template:6

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:7

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

mamba_ssm/ops/triton/mamba3/grouped_head_reduction.py· 2

Triton dependency

import triton

template:6

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:7

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

mamba_ssm/ops/triton/mamba3/mamba3_mimo_rotary_step.py· 5

Triton dependency

import triton

template:10

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:11

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.testing

template:12

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(q.device.index):

template:203

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

device = "cuda"

template:330

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/ops/triton/mamba3/mamba3_mimo_utils.py· 40

Triton dependency

import triton

template:49

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:50

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Device string "cuda"

dSSdA = torch.randn(B, H, nchunks, C, C, device='cuda', dtype=torch.float32)

template:1426

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

ddA_cs = torch.randn(B, H, S, device='cuda', dtype=torch.float32)

template:1427

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

ddA_cs_rev = torch.randn(B, H, S, device='cuda', dtype=torch.float32)

template:1428

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dA_cs = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1 # Scale to avoid overflow

template:1429

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dA_cs_rev = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1430

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

trap = torch.rand(B, H, S, device='cuda', dtype=torch.float16)

template:1474

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dt = F.softplus(-3.0 + torch.randn(B, H, S, device='cuda', dtype=torch.float))

template:1475

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dfactor = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1476

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dgamma_diag = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1477

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

da = -F.softplus(-3.0 + torch.randn(B, H, S, device='cuda', dtype=torch.float))

template:1518

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device='cuda',

template:1521

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

da = -F.softplus(-3.0 + torch.randn(B, H, S, device='cuda', dtype=torch.float))

template:1560

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device='cuda', dtype=torch.int32,

template:1601

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dSSdA = torch.randn(B, H, nchunks_global, C, C, device='cuda', dtype=torch.float32)

template:1610

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

ddA_cs = torch.randn(B, H, S, device='cuda', dtype=torch.float32)

template:1611

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

ddA_cs_rev = torch.randn(B, H, S, device='cuda', dtype=torch.float32)

template:1612

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dA_cs = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1613

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dA_cs_rev = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1614

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

SSdA = torch.zeros(B, H, nchunks_global, C, C, device='cuda', dtype=torch.float32)

template:1618

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dA_cs_padded = torch.zeros(B, H, curr_padded_len, device='cuda', dtype=torch.float32)

template:1627

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device='cuda', dtype=torch.int32,

template:1667

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

trap = torch.rand(B, H, S, device='cuda', dtype=torch.float16)

template:1671

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dt = F.softplus(-3.0 + torch.randn(B, H, S, device='cuda', dtype=torch.float32))

template:1672

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dfactor = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1673

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dgamma_diag = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1674

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Triton dependency

from triton.testing import do_bench

template:1706

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Device string "cuda"

dSSdA = torch.randn(B, H, nchunks, C, C, device='cuda', dtype=torch.float32)

template:1723

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

ddA_cs = torch.randn(B, H, S, device='cuda', dtype=torch.float32)

template:1724

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

ddA_cs_rev = torch.randn(B, H, S, device='cuda', dtype=torch.float32)

template:1725

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dA_cs = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1726

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dA_cs_rev = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1727

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Triton dependency

from triton.testing import do_bench

template:1769

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Device string "cuda"

da = F.softplus(-3.0 + torch.randn(B, H, S, device='cuda', dtype=torch.float))

template:1783

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Triton dependency

from triton.testing import do_bench

template:1806

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Device string "cuda"

trap = torch.ones(B, H, S, device='cuda', dtype=torch.float16) * 0.5

template:1820

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dt = F.softplus(-3.0 + torch.randn(B, H, S, device='cuda', dtype=torch.float))

template:1821

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dfactor = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1822

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dgamma_diag = torch.randn(B, H, S, device='cuda', dtype=torch.float32) * 0.1

template:1823

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/ops/triton/mamba3/mamba3_siso_bwd.py· 3

Triton dependency

import triton

template:14

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:15

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Device string "cuda"

return torch.empty(size, device="cuda", dtype=torch.int8)

template:1785

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/ops/triton/mamba3/mamba3_siso_combined.py· 2

Triton dependency

import triton

template:13

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Device string "cuda"

return torch.empty(size, device="cuda", dtype=torch.int8)

template:23

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/ops/triton/mamba3/mamba3_siso_fwd.py· 3

Triton dependency

import triton

template:14

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:15

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Device string "cuda"

return torch.empty(size, device="cuda", dtype=torch.int8)

template:437

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/ops/triton/mamba3/mamba3_siso_step.py· 3

Triton dependency

import triton

template:12

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:13

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Device string "cuda"

return torch.empty(size, device="cuda", dtype=torch.int8)

template:230

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/ops/triton/mamba3/utils.py· 2

Triton dependency

import triton

template:7

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:8

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

mamba_ssm/ops/triton/selective_state_update.py· 3

Triton dependency

import triton

template:10

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:11

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:200

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/ops/triton/softplus.py· 2

Triton dependency

import triton

template:1

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:2

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

mamba_ssm/ops/triton/ssd_bmm.py· 4

Triton dependency

import triton

template:10

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:11

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(a.device.index):

template:198

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(a.device.index):

template:253

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/ops/triton/ssd_chunk_scan.py· 13

Triton dependency

import triton

template:12

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:13

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:1390

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(C.device.index):

template:1436

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

sm_count = torch.cuda.get_device_properties(dout.device).multi_processor_count

template:1485

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(dout.device.index):

template:1491

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

sm_count = torch.cuda.get_device_properties(x.device).multi_processor_count

template:1534

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:1540

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:1592

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:1640

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:1680

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:1713

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(dout.device.index):

template:1747

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/ops/triton/ssd_chunk_state.py· 10

Triton dependency

import triton

template:10

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:11

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(dt.device.index):

template:727

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(dt.device.index):

template:787

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:829

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:878

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

sm_count = torch.cuda.get_device_properties(x.device).multi_processor_count

template:937

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:943

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:996

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:1032

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/ops/triton/ssd_combined.py· 3

Triton dependency

import triton

template:16

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:17

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(x.device.index):

template:308

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/ops/triton/ssd_state_passing.py· 4

Triton dependency

import triton

template:10

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

Triton dependency

import triton.language as tl

template:11

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

with torch.cuda.device(states.device.index):

template:210

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.device(dout.device.index):

template:258

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/utils/determinism.py· 1

Triton dependency

import triton

template:10

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

pyproject.toml· 1

Triton dependency

"triton>=3.5.0",

template:23

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

tests/benchmark_determinism_kernels.py· 10

Triton dependency

from triton.testing import do_bench

template:8

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

torch.cuda.empty_cache()

template:20

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.reset_peak_memory_stats()

template:21

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.synchronize()

template:22

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.synchronize()

template:28

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.synchronize()

template:31

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

return torch.cuda.max_memory_allocated() / (1024 * 1024)

template:32

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

device = "cuda"

template:37

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

if not torch.cuda.is_available():

template:103

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

print(f"GPU: {torch.cuda.get_device_name(0)}")

template:118

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

tests/ops/cute/test_mamba3_mimo_step.py· 8

Device string "cuda"

DEVICE = "cuda"

template:39

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

if not torch.cuda.is_available():

template:56

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.manual_seed_all(42)

template:161

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.manual_seed_all(42)

template:167

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.manual_seed_all(42)

template:176

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.manual_seed_all(42)

template:182

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Triton dependency

from triton.testing import do_bench_cudagraph

template:265

OpenAI Triton has an AMD backend. Kernels usually run on ROCm with a matching Triton build; occasional tuning is needed.

RecommendInstall the ROCm-enabled Triton build.

torch.cuda API usage

torch.cuda.manual_seed_all(42)

template:272

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

benchmarks/benchmark_generation_mamba_simple.py· 5

Device string "cuda"

device = "cuda"

template:31

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

input_ids = torch.randint(1, 1000, (args.batch, args.promptlen), dtype=torch.long, device="cuda")

template:47

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

attn_mask = torch.ones_like(input_ids, dtype=torch.long, device="cuda")

template:48

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

torch.cuda.synchronize()

template:86

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.synchronize()

template:90

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

evals/lm_harness_eval.py· 1

Device string "cuda"

def __init__(self, pretrained="state-spaces/mamba-2.8b", max_length=2048, batch_size=None, device="cuda",

template:19

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/modules/ssd_minimal.py· 1

Device string "cuda"

device = "cuda"

template:92

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/ops/tilelang/mamba3/mamba3_mimo_fwd_varlen.py· 6

Device string "cuda"

o = torch.empty((B, S, H, P), device='cuda', dtype=dtype)

template:693

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

o = torch.empty((B, S, R, H, P), device='cuda', dtype=dtype)

template:695

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

h = torch.empty((NS, H, N, P), device='cuda', dtype=torch.float32) if return_state else None

template:723

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

k_final = torch.empty((NS, R, H, N), device='cuda', dtype=dtype) if return_state else None

template:724

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

h = torch.empty((B, H, N, P), device='cuda', dtype=torch.float32) if return_state else None

template:726

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

k_final = torch.empty((B, R, H, N), device='cuda', dtype=dtype) if return_state else None

template:727

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/ops/tilelang/mamba3/mamba3_mimo_fwd.py· 4

Device string "cuda"

o = torch.empty((B, S, H, P), device='cuda', dtype=dtype)

template:515

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

o = torch.empty((B, S, R, H, P), device='cuda', dtype=dtype)

template:517

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

h = torch.empty((B, H, N, P), device='cuda', dtype=torch.float32) if return_state else None

template:544

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

k_final = torch.empty((B, R, H, N), device='cuda', dtype=dtype) if return_state else None

template:545

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

mamba_ssm/utils/generation.py· 10

torch.cuda API usage

start = torch.cuda.Event(enable_timing=enable_timing)

template:214

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

end = torch.cuda.Event(enable_timing=enable_timing)

template:215

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.synchronize()

template:241

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

cache.mempool = torch.cuda.graphs.graph_pool_handle()

template:320

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

s = torch.cuda.Stream()

template:353

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

s.wait_stream(torch.cuda.current_stream())

template:354

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.stream(s):

template:355

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.current_stream().wait_stream(s)

template:369

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

graph = torch.cuda.CUDAGraph()

template:372

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

with torch.cuda.graph(graph, pool=mempool):

template:373

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

mamba_ssm/utils/torch.py· 1

Device string "cuda"

kwargs["device_type"] = "cuda"

template:8

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

tests/modules/test_mamba3_varlen.py· 4

torch.cuda API usage

if not torch.cuda.is_available():

template:17

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

device="cuda",

template:36

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = "cuda"

template:68

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = "cuda"

template:103

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

tests/ops/test_selective_scan.py· 2

Device string "cuda"

device = 'cuda'

template:42

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = 'cuda'

template:161

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

tests/ops/tilelang/test_mamba3_mimo.py· 47

torch.cuda API usage

if not torch.cuda.is_available():

template:72

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

torch.cuda.manual_seed_all(seed)

template:150

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

q = torch.randn((b, s, r, g, n), device="cuda", dtype=dtype)

template:152

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

k = torch.randn((b, s, r, g, n), device="cuda", dtype=dtype)

template:153

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

v = torch.randn((b, s, h, p), device="cuda", dtype=dtype)

template:154

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

q_bias = torch.randn((h, r, n), device="cuda", dtype=torch.float32)

template:156

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

k_bias = torch.randn((h, r, n), device="cuda", dtype=torch.float32)

template:157

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

mimo_v = torch.randn((h, r, p), device="cuda", dtype=torch.float32) / r

template:158

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

mimo_o = torch.randn((h, r, p), device="cuda", dtype=torch.float32) / r

template:159

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

d = torch.randn((h,), device="cuda", dtype=torch.float32) if has_d else None

template:163

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

(b, s, h, n // rotary_dim_divisor), device="cuda", dtype=torch.float32

template:166

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dt = F.softplus(-3.0 + torch.randn((b, h, s), device="cuda", dtype=torch.float32))

template:168

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

a = torch.rand((b, h, s), device="cuda", dtype=torch.float32)

template:169

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

trap = torch.rand((b, h, s), device="cuda", dtype=dtype)

template:172

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

outproj_norm_weight = torch.randn((h, p), device="cuda", dtype=torch.float32)

template:174

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device: str = "cuda",

template:211

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

if device == "cuda":

template:216

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

torch.cuda.manual_seed_all(seed)

template:217

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

dout_prereduce = torch.randn((b, s, r, h, p_dim), device="cuda", dtype=FIXED_DTYPE)

template:1363

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device="cuda",

template:1607

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device="cuda", dtype=torch.int32,

template:1660

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

torch.cuda.manual_seed_all(42)

template:1664

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

(1, s_total, mimo_rank, nheads_qk, headdim_qk), device="cuda",

template:1667

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

(1, s_total, nheads, headdim_v), device="cuda",

template:1672

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

-3.0 + torch.randn(1, nheads, s_total, device="cuda", dtype=torch.float)

template:1676

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

torch.rand((1, nheads, s_total), device="cuda", dtype=FIXED_DTYPE) * 0.5

template:1680

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

(nheads, mimo_rank, headdim_qk), device="cuda",

template:1683

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

torch.randn((nheads, mimo_rank, headdim_v), device="cuda", dtype=torch.float32) / mimo_rank

template:1688

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

(1, s_total, nheads, headdim_qk // rotary_dim_divisor), device="cuda",

template:1693

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

D = torch.randn((nheads,), device="cuda", dtype=torch.float32, requires_grad=True)

template:1696

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

(1, s_total, nheads, headdim_v), device="cuda",

template:1698

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

torch.cuda.manual_seed_all(seed)

template:1789

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

device="cuda",

template:1797

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

q = torch.randn((b, s_total, r, g, n), device="cuda", dtype=dtype)

template:1801

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

k = torch.randn((b, s_total, r, g, n), device="cuda", dtype=dtype)

template:1802

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

v = torch.randn((b, s_total, h, p), device="cuda", dtype=dtype)

template:1803

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

q_bias = torch.randn((h, r, n), device="cuda", dtype=torch.float32)

template:1805

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

k_bias = torch.randn((h, r, n), device="cuda", dtype=torch.float32)

template:1806

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

mimo_v = torch.randn((h, r, p), device="cuda", dtype=torch.float32) / r

template:1807

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

mimo_o = torch.randn((h, r, p), device="cuda", dtype=torch.float32) / r

template:1808

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

d = torch.randn((h,), device="cuda", dtype=torch.float32) if has_d else None

template:1812

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

(b, s_total, h, n // rotary_dim_divisor), device="cuda", dtype=torch.float32

template:1815

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

dt = F.softplus(-3.0 + torch.randn((b, h, s_total), device="cuda", dtype=torch.float32))

template:1817

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

a = torch.rand((b, h, s_total), device="cuda", dtype=torch.float32)

template:1818

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

trap = torch.rand((b, h, s_total), device="cuda", dtype=dtype)

template:1825

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

outproj_norm_weight = torch.randn((h, p), device="cuda", dtype=torch.float32)

template:1827

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

(b, s_total, r, h_dim, p_dim), device="cuda", dtype=FIXED_DTYPE

template:2349

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

tests/ops/triton/test_layernorm_gated.py· 1

Device string "cuda"

device = 'cuda'

template:34

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

tests/ops/triton/test_mamba3_siso.py· 4

Device string "cuda"

device = 'cuda'

template:489

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = 'cuda'

template:564

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = 'cuda'

template:678

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = 'cuda'

template:845

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

tests/ops/triton/test_selective_state_update.py· 4

Device string "cuda"

device = "cuda"

template:23

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = "cuda"

template:67

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = "cuda"

template:113

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = "cuda"

template:162

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

tests/ops/triton/test_ssd.py· 1

Device string "cuda"

device = 'cuda'

template:31

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

tests/test_determinism.py· 16

torch.cuda API usage

torch.cuda.manual_seed_all(seed)

template:17

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

device = "cuda"

template:37

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

torch.cuda.synchronize()

template:139

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")

template:182

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")

template:190

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")

template:198

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")

template:206

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")

template:214

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

device = "cuda"

template:218

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

torch.cuda.synchronize()

template:238

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")

template:264

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

device = "cuda"

template:268

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

torch.cuda.synchronize()

template:290

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

torch.cuda API usage

@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")

template:307

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

Device string "cuda"

device = "cuda"

template:311

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

torch.cuda API usage

torch.cuda.synchronize()

template:329

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

RecommendNo change required. Install the ROCm build of PyTorch (pip install torch --index-url https://download.pytorch.org/whl/rocm6.1).

tests/test_generation.py· 3

Device string "cuda"

device = "cuda"

template:16

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = "cuda"

template:50

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.

Device string "cuda"

device = "cuda"

template:123

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

RecommendNo change required on ROCm PyTorch.