[Draft][Quantization][MoE] NVFP4 native expert kernels (raw-layout decode GEMV + grouped prefill) - #44

Draft

01554 wants to merge 16 commits into

lab/upstream-base-20260909from

lab/upstream-native-nvfp4-backend

Commits on Sep 9, 2026

[Quantization][MoE] NVFP4 native expert kernels: raw-layout decode GEMV and grouped prefill (kernels + tests only)
01554
committed
1baf5f7
[Quantization][MoE] nvfp4_native tests: normal imports with an environment skip, no file-location shim
01554
committed
9c78017
[Quantization][MoE] NVFP4 native backend: explicit moe_backend=native selection and device-resident experts
01554
committed
59b98f0
[Quantization][MoE] nvfp4_native experts: pass router weights to the adapters as FP32 (review)
01554
committed
9179e6a
[Quantization][MoE] nvfp4_native experts: share decode/prefill scratch across layers of the same shape (review)
01554
committed
b972f3d
[Quantization][MoE] nvfp4_native loader: docstring describes moe_backend=native selection (review)
01554
committed
cb90421
[Quantization][MoE] nvfp4_native experts: reject non-BF16 activations, LoRA, bias and SwiGLU parameters explicitly (review)
01554
committed
aa83ec6
[Quantization][MoE] nvfp4_native loader test: the loader leaves the method's kernel untouched (review)
01554
committed
f8951a3
[Quantization][MoE] nvfp4_native experts: scratch shared per (device, stream, shape), resolved on the forward stream (review)
01554
committed
a232b4d
[Quantization][MoE] nvfp4_native experts: all large scratch from the modular kernel workspace (review)
01554
committed
61b5343
[Quantization][MoE] nvfp4_native experts: decode carve uses the actual row count (review)
01554
committed
b5a302c
[Quantization][MoE] nvfp4_native experts test: assert the exact unsupported-reason words (review)
01554
committed
becb832
[Quantization][MoE] nvfp4_native tests: drop the remaining importorskip guard (full-environment failures must fail)
01554
committed
a452302
[Quantization][MoE] nvfp4_native: exercise the real FusedMoEKernel + WorkspaceManager path (CPU)
01554
committed
74debf5
tests: pass the required expert_map / apply_router_weight_on_input to FusedMoEKernel.apply
01554
committed
75686ad
tests: pass topk_weights/topk_ids to FusedMoEKernel.apply by keyword (order was swapped)
01554
committed
848f0ef

Navigation Menu

[Draft][Quantization][MoE] NVFP4 native expert kernels (raw-layout decode GEMV + grouped prefill) - #44

[Draft][Quantization][MoE] NVFP4 native expert kernels (raw-layout decode GEMV + grouped prefill)#44
01554 wants to merge 16 commits into
lab/upstream-base-20260909from
lab/upstream-native-nvfp4-backend

Commits on Sep 9, 2026

[Quantization][MoE] NVFP4 native expert kernels: raw-layout decode GEMV and grouped prefill (kernels + tests only)

[Quantization][MoE] nvfp4_native tests: normal imports with an environment skip, no file-location shim

[Quantization][MoE] NVFP4 native backend: explicit moe_backend=native selection and device-resident experts

[Quantization][MoE] nvfp4_native experts: pass router weights to the adapters as FP32 (review)

[Quantization][MoE] nvfp4_native experts: share decode/prefill scratch across layers of the same shape (review)

[Quantization][MoE] nvfp4_native loader: docstring describes moe_backend=native selection (review)

[Quantization][MoE] nvfp4_native experts: reject non-BF16 activations, LoRA, bias and SwiGLU parameters explicitly (review)

[Quantization][MoE] nvfp4_native loader test: the loader leaves the method's kernel untouched (review)

[Quantization][MoE] nvfp4_native experts: scratch shared per (device, stream, shape), resolved on the forward stream (review)

[Quantization][MoE] nvfp4_native experts: all large scratch from the modular kernel workspace (review)

[Quantization][MoE] nvfp4_native experts: decode carve uses the actual row count (review)

[Quantization][MoE] nvfp4_native experts test: assert the exact unsupported-reason words (review)

[Quantization][MoE] nvfp4_native tests: drop the remaining importorskip guard (full-environment failures must fail)

[Quantization][MoE] nvfp4_native: exercise the real FusedMoEKernel + WorkspaceManager path (CPU)

tests: pass the required expert_map / apply_router_weight_on_input to FusedMoEKernel.apply

tests: pass topk_weights/topk_ids to FusedMoEKernel.apply by keyword (order was swapped)