Skip to content

[Atlas Inference] Strix Halo (gfx1151) Mamba-2 + conv1d Hub kernel path #49709

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
AzeezIsh wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[Atlas Inference] Strix Halo (gfx1151) Mamba-2 + conv1d Hub kernel path #49709

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions docs/source/en/model_doc/nemotron_h.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,13 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
## Notes

- Use left padding for batched generation. See [Padding side](../llm_tutorial#padding-side).
- On NVIDIA GB10 (`SM121`) and AMD Strix Halo (`gfx1151`), `mamba-ssm` and `causal-conv1d` have no prebuilt wheels, so `NemotronHMamba2Mixer` falls back to the slow PyTorch `einsum` segment-sum reference. Pass `use_kernels=True` (`pip install -U kernels`) to [`~PreTrainedModel.from_pretrained`] to use the Hub kernels instead (`kernels-community/mamba-ssm` `v4` on CUDA and [`Atlas-Inference/gdn`](https://huggingface.co/kernels/Atlas-Inference/gdn) on `gfx1151` ROCm). Measured on Strix Halo (`gfx1151`, torch 2.14 + ROCm 7.2, `Nemotron-3.5-Super-VL-120B` Mamba-2 mixer dimensions, bf16):

| Prompt Tokens | `use_kernels=False` TTFT | `use_kernels=True` TTFT | Decode (`False` -> `True`) |
| --- | --- | --- | --- |
| 1,024 | 0.64 s | 0.38 s (1.67x faster) | 27.8 -> 31.5 tok/s |
| 4,096 | 2.71 s | 1.65 s (1.64x faster) | 24.7 -> 29.2 tok/s |
| 16,384 | 18.32 s | 9.38 s (1.95x faster) | 20.9 -> 23.0 tok/s |

## NemotronHConfig

Expand Down
118 changes: 89 additions & 29 deletions src/transformers/integrations/hub_kernels.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -178,7 +178,7 @@ def _build_kernel_mapping() -> dict:
): LayerRepository(
repo_id="Atlas-Inference/gdn",
layer_name="Qwen3_5GatedDeltaNet",
revision="13462bf5e936315f341204ae06bf09068392d227",
revision="f9704654e7c18d3149dfa314fed44eaee52f6ca9",
# TODO: drop once Atlas-Inference is an allow-listed trusted publisher
trust_remote_code=True,
),
Expand All @@ -188,24 +188,36 @@ def _build_kernel_mapping() -> dict:
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="causal_conv1d_fn",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="causal_conv1d_fn",
version=3,
version=4,
),
},
Device(
type="rocm",
properties=ROCMProperties(min_capability=115, max_capability=115),
): {
Mode.INFERENCE: LayerRepository(
repo_id="Atlas-Inference/gdn",
layer_name="causal_conv1d_fn",
revision="f9704654e7c18d3149dfa314fed44eaee52f6ca9",
# TODO: drop once Atlas-Inference is an allow-listed trusted publisher
trust_remote_code=True,
),
},
"xpu": {
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="causal_conv1d_fn",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="causal_conv1d_fn",
version=3,
version=4,
),
},
},
Expand All @@ -214,24 +226,36 @@ def _build_kernel_mapping() -> dict:
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="causal_conv1d_update",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="causal_conv1d_update",
version=3,
version=4,
),
},
Device(
type="rocm",
properties=ROCMProperties(min_capability=115, max_capability=115),
): {
Mode.INFERENCE: LayerRepository(
repo_id="Atlas-Inference/gdn",
layer_name="causal_conv1d_update",
revision="f9704654e7c18d3149dfa314fed44eaee52f6ca9",
# TODO: drop once Atlas-Inference is an allow-listed trusted publisher
trust_remote_code=True,
),
},
"xpu": {
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="causal_conv1d_update",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="causal_conv1d_update",
version=3,
version=4,
),
},
},
Expand Down Expand Up @@ -284,24 +308,36 @@ def _build_kernel_mapping() -> dict:
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_chunk_scan_combined",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_chunk_scan_combined",
version=3,
version=4,
),
},
Device(
type="rocm",
properties=ROCMProperties(min_capability=115, max_capability=115),
): {
Mode.INFERENCE: LayerRepository(
repo_id="Atlas-Inference/gdn",
layer_name="mamba_chunk_scan_combined",
revision="f9704654e7c18d3149dfa314fed44eaee52f6ca9",
# TODO: drop once Atlas-Inference is an allow-listed trusted publisher
trust_remote_code=True,
),
},
"xpu": {
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_chunk_scan_combined",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_chunk_scan_combined",
version=3,
version=4,
),
},
},
Expand All @@ -310,24 +346,36 @@ def _build_kernel_mapping() -> dict:
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_split_conv1d_scan_combined",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_split_conv1d_scan_combined",
version=3,
version=4,
),
},
Device(
type="rocm",
properties=ROCMProperties(min_capability=115, max_capability=115),
): {
Mode.INFERENCE: LayerRepository(
repo_id="Atlas-Inference/gdn",
layer_name="mamba_split_conv1d_scan_combined",
revision="f9704654e7c18d3149dfa314fed44eaee52f6ca9",
# TODO: drop once Atlas-Inference is an allow-listed trusted publisher
trust_remote_code=True,
),
},
"xpu": {
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_split_conv1d_scan_combined",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_split_conv1d_scan_combined",
version=3,
version=4,
),
},
},
Expand All @@ -336,24 +384,24 @@ def _build_kernel_mapping() -> dict:
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_inner_fn",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_inner_fn",
version=3,
version=4,
),
},
"xpu": {
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_inner_fn",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="mamba_inner_fn",
version=3,
version=4,
),
},
},
Expand All @@ -362,24 +410,24 @@ def _build_kernel_mapping() -> dict:
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="selective_scan_fn",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="selective_scan_fn",
version=3,
version=4,
),
},
"xpu": {
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="selective_scan_fn",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="selective_scan_fn",
version=3,
version=4,
),
},
},
Expand All @@ -388,24 +436,36 @@ def _build_kernel_mapping() -> dict:
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="selective_state_update",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="selective_state_update",
version=3,
version=4,
),
},
Device(
type="rocm",
properties=ROCMProperties(min_capability=115, max_capability=115),
): {
Mode.INFERENCE: LayerRepository(
repo_id="Atlas-Inference/gdn",
layer_name="selective_state_update",
revision="f9704654e7c18d3149dfa314fed44eaee52f6ca9",
# TODO: drop once Atlas-Inference is an allow-listed trusted publisher
trust_remote_code=True,
),
},
"xpu": {
Mode.TRAINING: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="selective_state_update",
version=3,
version=4,
),
Mode.INFERENCE: LayerRepository(
repo_id="kernels-community/mamba-ssm",
layer_name="selective_state_update",
version=3,
version=4,
),
},
},
Expand Down
Loading