Module quantize 

Quantization op family — Category P.

Phase 8 splits across two parallel milestones:

• Milestone 8.1 (sibling): per-tensor + per-channel quantize / dequantize plus fake_quantize. Owns crates/baracuda-kernels-sys/kernels/quantize/per_tensor.cu / per_channel.cu / fake_quantize.cu and the Rust plans for those ops in this quantize/ module.

• Milestone 8.2 (this work): per-token + per-group quantize / dequantize plus their STE backwards. Used by LLM activation (W8A8 per-row) and weight (GPTQ-style INT4 per-group, g=128) quant. Owns crates/baracuda-kernels-sys/kernels/quantize/per_token.cu / per_group.cu and the plans in this module.

The two milestones share append-only edits to this file, to crate::lib’s re-exports, and to baracuda-kernels-sys/src/lib.rs. No existing entry is rewritten.

Trailblazer dtype coverage: input FP ∈ {f32, f64, f16, bf16}; output int ∈ {s8, u8}. Sub-byte packed types (s4 / u4) are deferred to a later milestone.

Backward convention is the Straight-Through Estimator (STE): dx = (dy / scale) * 1[qmin < q < qmax]. The in-range mask is recomputed inside the BW kernel from the saved input — callers must retain the input tensor for autograd (which they would do anyway).

Re-exports

pub use dequantize_per_channel::DequantizePerChannelArgs;

pub use dequantize_per_channel::DequantizePerChannelDescriptor;

pub use dequantize_per_channel::DequantizePerChannelPlan;

pub use dequantize_per_channel_backward::DequantizePerChannelBackwardArgs;

pub use dequantize_per_channel_backward::DequantizePerChannelBackwardDescriptor;

pub use dequantize_per_channel_backward::DequantizePerChannelBackwardPlan;

pub use dequantize_per_tensor::DequantizePerTensorArgs;

pub use dequantize_per_tensor::DequantizePerTensorDescriptor;

pub use dequantize_per_tensor::DequantizePerTensorPlan;

pub use dequantize_per_tensor_backward::DequantizePerTensorBackwardArgs;

pub use dequantize_per_tensor_backward::DequantizePerTensorBackwardDescriptor;

pub use dequantize_per_tensor_backward::DequantizePerTensorBackwardPlan;

pub use fake_quantize::FakeQuantizeArgs;

pub use fake_quantize::FakeQuantizeDescriptor;

pub use fake_quantize::FakeQuantizePlan;

pub use fake_quantize_backward::FakeQuantizeBackwardArgs;

pub use fake_quantize_backward::FakeQuantizeBackwardDescriptor;

pub use fake_quantize_backward::FakeQuantizeBackwardPlan;

pub use per_channel::QuantizePerChannelArgs;

pub use per_channel::QuantizePerChannelDescriptor;

pub use per_channel::QuantizePerChannelPlan;

pub use per_channel_backward::QuantizePerChannelBackwardArgs;

pub use per_channel_backward::QuantizePerChannelBackwardDescriptor;

pub use per_channel_backward::QuantizePerChannelBackwardPlan;

pub use per_tensor::QuantizePerTensorArgs;

pub use per_tensor::QuantizePerTensorDescriptor;

pub use per_tensor::QuantizePerTensorPlan;

pub use per_tensor_backward::QuantizePerTensorBackwardArgs;

pub use per_tensor_backward::QuantizePerTensorBackwardDescriptor;

pub use per_tensor_backward::QuantizePerTensorBackwardPlan;

pub use per_token::QuantizePerTokenArgs;

pub use per_token::QuantizePerTokenDescriptor;

pub use per_token::QuantizePerTokenPlan;

pub use per_token_backward::QuantizePerTokenBackwardArgs;

pub use per_token_backward::QuantizePerTokenBackwardDescriptor;

pub use per_token_backward::QuantizePerTokenBackwardPlan;

pub use dequantize_per_token::DequantizePerTokenArgs;

pub use dequantize_per_token::DequantizePerTokenDescriptor;

pub use dequantize_per_token::DequantizePerTokenPlan;

pub use dequantize_per_token_backward::DequantizePerTokenBackwardArgs;

pub use dequantize_per_token_backward::DequantizePerTokenBackwardDescriptor;

pub use dequantize_per_token_backward::DequantizePerTokenBackwardPlan;

pub use per_group::QuantizePerGroupArgs;

pub use per_group::QuantizePerGroupDescriptor;

pub use per_group::QuantizePerGroupPlan;

pub use per_group_backward::QuantizePerGroupBackwardArgs;

pub use per_group_backward::QuantizePerGroupBackwardDescriptor;

pub use per_group_backward::QuantizePerGroupBackwardPlan;

pub use dequantize_per_group::DequantizePerGroupArgs;

pub use dequantize_per_group::DequantizePerGroupDescriptor;

pub use dequantize_per_group::DequantizePerGroupPlan;

pub use dequantize_per_group_backward::DequantizePerGroupBackwardArgs;

pub use dequantize_per_group_backward::DequantizePerGroupBackwardDescriptor;

pub use dequantize_per_group_backward::DequantizePerGroupBackwardPlan;

pub use dynamic_range::DynamicRangeMode;

pub use dynamic_range::DynamicRangeQuantizeArgs;

pub use dynamic_range::DynamicRangeQuantizeDescriptor;

pub use dynamic_range::DynamicRangeQuantizePlan;

pub use dynamic_range::DynamicRangeScope;

pub use quantized_linear::QuantizedLinearArgs;

pub use quantized_linear::QuantizedLinearDescriptor;

pub use quantized_linear::QuantizedLinearPlan;

pub use smoothquant::SmoothQuantLinearArgs;

pub use smoothquant::SmoothQuantLinearDescriptor;

pub use smoothquant::SmoothQuantLinearPlan;

pub use gguf::BlockQ2K;

pub use gguf::BlockQ3K;

pub use gguf::BlockQ4_0;

pub use gguf::BlockQ4_1;

pub use gguf::BlockQ4K;

pub use gguf::BlockQ5_0;

pub use gguf::BlockQ5_1;

pub use gguf::BlockQ5K;

pub use gguf::BlockQ6K;

pub use gguf::BlockQ8_0;

pub use gguf::BlockQ8K;

pub use gguf::GgufDequantizeArgs;

pub use gguf::GgufDequantizeDescriptor;

pub use gguf::GgufDequantizePlan;

pub use gguf::GgufMmvqArgs;

pub use gguf::GgufMmvqDescriptor;

pub use gguf::GgufMmvqPlan;

pub use gguf::GgufMmvqBatchedActivation;

pub use gguf::GgufMmvqBatchedArgs;

pub use gguf::GgufMmvqBatchedDescriptor;

pub use gguf::GgufMmvqBatchedFormat;

pub use gguf::GgufMmvqBatchedPlan;

pub use gguf::GgufMmvqMultiMArgs;

pub use gguf::GgufMmvqMultiMDescriptor;

pub use gguf::GgufMmvqMultiMPlan;

pub use nf4::Nf4Activation;

pub use nf4::Nf4DequantizeArgs;

pub use nf4::Nf4DequantizePlan;

pub use nf4::Nf4Descriptor;

pub use nf4::Nf4MmvqArgs;

pub use nf4::Nf4MmvqMultiMArgs;

pub use nf4::Nf4MmvqMultiMDescriptor;

pub use nf4::Nf4MmvqMultiMPlan;

pub use nf4::Nf4MmvqPlan;

pub use nf4::NF4_CODEBOOK;

Modules

dequantize_per_channel

dequantize_per_channel forward plan.

dequantize_per_channel_backward

dequantize_per_channel backward plan — dq[i] = dy[i] * scale[c].

dequantize_per_group

dequantize_per_group forward plan.

dequantize_per_group_backward

dequantize_per_group backward plan — straight-through.

dequantize_per_tensor

dequantize_per_tensor forward plan.

dequantize_per_tensor_backward

dequantize_per_tensor backward plan — dq = dy * scale.

dequantize_per_token

dequantize_per_token forward plan.

dequantize_per_token_backward

dequantize_per_token backward plan — straight-through (dq = dy * scale[n]).

dynamic_range

dynamic_range_quantize — compose op (Phase 8 Milestone 8.3).

fake_quantize

fake_quantize forward plan — per-tensor, FP roundtrip.

fake_quantize_backward

fake_quantize backward plan via STE.

gguf

GGUF block-format quantization plans — Phase 8 Milestone 8.4.

nf4

NF4 (NormalFloat 4-bit) dequant + GEMV plans — Phase 53.

per_channel

quantize_per_channel forward plan.

per_channel_backward

quantize_per_channel backward plan via STE.

per_group

quantize_per_group forward plan.

per_group_backward

quantize_per_group backward plan (Straight-Through Estimator).

per_tensor

quantize_per_tensor forward plan — Category P FW trailblazer.

per_tensor_backward

quantize_per_tensor backward plan (Straight-Through Estimator).

per_token

quantize_per_token forward plan.

per_token_backward

quantize_per_token backward plan (Straight-Through Estimator).

quantized_linear

quantized_linear — fused W8A8 quantized matmul (Phase 8.3).

smoothquant

SmoothQuantLinearPlan — Phase 45 zero-new-CUDA composition.

Functions

default_q_range

Default qmin / qmax for an output integer dtype. Today wired for the two trailblazer output kinds — baracuda_kernels_types::S8 ([-128, 127]) and baracuda_kernels_types::U8 ([0, 255]).