[ET-VK][Ops] torchao.choose_qparams_affine vulkan impl and shader (buffer only) and cleanup #13003
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
added 4 commits
July 29, 2025 23:43
Pull Request resolved: #12574 # Changes * Introduce `linear_qta8a_qga4w` custom operator in `custom_ops_lib.py` to handle dynamic activation + grouped weight quantized linear operations * Add pattern matching and fusion logic in `FuseQuantizedOpsTransform` to detect and replace dequant + dequant + linear sequences with the new fused operator * Implement comprehensive test coverage in `test_vulkan_passes.py` for the QTA8A_QGA4W fusion pattern validation * Add 4-bit weight packing utilities and grouped quantization support for efficient memory usage # Motivation The existing quantization workflow in Vulkan backend processes dynamic activation + grouped weight quantized linear operations as separate quantize/dequantize/linear steps, which creates performance overhead through: * Multiple kernel dispatches instead of a single fused operation * Intermediate tensor allocations for dequantized weights and activations * Suboptimal memory bandwidth utilization The new `linear_qta8a_qga4w` operator fuses the entire sequence into a single operation that: * Directly processes 8-bit quantized activations with per-token scales/zero-points * Handles 4-bit grouped quantized weights with configurable group sizes * Eliminates intermediate dequantization steps by performing dequantization inline * Reduces memory footprint through packed 4-bit weight storage This aligns with the broader goal of optimizing quantized model inference in the Vulkan backend by leveraging graph-level transformations to improve computational efficiency while maintaining numerical accuracy. ghstack-source-id: 299473616 @exported-using-ghexport Differential Revision: [D78291269](https://our.internmc.facebook.com/intern/diff/D78291269/)
Pull Request resolved: #12575 # Changes * Implement `torchao.quantize_affine` operator in Vulkan backend with comprehensive texture and buffer storage support * Add block-wise quantization mode in `quantize_texture.glsl` and `quantize_buffer.glsl` shaders for configurable tensor block quantization * Introduce comprehensive test suite in `affine_test.cpp` with multi-dimensional tensor validation and reference implementation * Extend quantization infrastructure in `Quantize.cpp` to handle affine transformations with configurable block sizes and quantization parameters BE: Improved the documentation in the shader logic which is more detailed and clear NOTE: I delegated the quantize_affine and future affine operators through a new custom test file denoted as `affine_test.cpp` as the other quantization testing framework was getting a little large, and it makes more sense to separate the namespace between torchao and quantized_decomposed. I believe the _decomposed namespace is getting phased out in favor of this affine operator so deprecation will be easier in the future. # Motivation The existing Vulkan quantization infrastructure lacked support for the `torchao.quantize_affine` operator, which is essential for enabling dynamic quantization efficiently. The `quantize_affine` operator provides flexible block-wise quantization that allows different scale and zero-point values for tensor blocks, enabling: * **Block-wise Quantization**: Applies quantization parameters to configurable tensor blocks rather than entire tensors, improving quantization accuracy for heterogeneous data distributions * **Affine Transformation**: Uses the formula `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)` for precise floating-point to integer mapping # Operator Description The `quantize_affine` operator converts floating-point tensor values to n-bit integer representations using pre-computed quantization parameters (scale and zero_point) applied to configurable tensor blocks. Block-wise quantization divides tensors into blocks and applies separate quantization parameters to each block, allowing fine-grained control over quantization precision. The quantization formula is: `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)` **Storage Requirements**: Scale and zero_point tensors must use buffer storage with width-packed layout. Input/output tensors support both buffer and texture storage with standard axis mapping. # Block-wise Quantization Implementation Block-wise quantization enables fine-grained quantization by dividing tensors into blocks and applying separate quantization parameters to each block. The implementation uses several key data structures computed in `Quantize.cpp`: * **`block_size_vec`**: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks) * **`tensor_size_whcn`**: Input tensor dimensions converted to WHCN layout using `utils::make_whcn_ivec4()` * **`num_blocks_vec`**: Number of blocks per dimension calculated as `tensor_size_whcn / block_size_vec` * **`block_stride_vec`**: Pre-computed linear strides for block grid indexing `{1, #W, #W*#H, #W*#H*#C}` to enable efficient block ID calculation The block coordinate calculation uses: `bcoord = tidx / blockSize` where `tidx` is the tensor coordinate in WHCN layout, then the linear block ID is computed as: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w` # Shader Algorithm Overview ## Texture Storage Implementation (`quantize_texture.glsl`) **Workgroup Configuration**: - **Global WG Size**: Default sizing based on texture dimensions - **Local WG Size**: Default with special handling for batch dimension quantization (Z dimension set to 1 for proper workgroup dispatching when `global_workgroup_size[2] > 1`) **Block-wise Mode Algorithm**: The shader processes 3D texture positions where each position represents a texel containing 4 width-packed components. For each texel at position `pos`, it calculates a base tensor index `base_tidx = ivec4(pos.x * 4, pos.y, pos.z, 0)` to account for width-packing. For each of the 4 components in the texel, it computes the actual tensor coordinate: `tidx = ivec4(base_tidx.x + i, base_tidx.y, (foldedZ % C_total), (foldedZ / C_total))` where `foldedZ = pos.z` handles batch-channel folding in 4D tensors and `C_total = numBlocks.z * blockSize.z` represents the total channel dimension. The block coordinate is calculated using integer division: `bcoord = tidx / blockSize`, then the linear block ID uses pre-computed strides: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`. Each component is quantized using its corresponding block's parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])` and written to the output texel. ## Buffer Storage Implementation (`quantize_buffer.glsl`) **Workgroup Configuration**: - **Global WG Size**: Default sizing based on buffer element count - **Local WG Size**: Default sizing without special constraints **Block-wise Mode Algorithm**: The shader processes linear buffer indices using `gl_GlobalInvocationID.x` as the output buffer index. It converts this to tensor coordinates using `bufi_to_tidx(out_bufi, t_out_strides, out_dim_order)` which handles the buffer-to-tensor index mapping with proper stride calculations. For each element, it computes the block coordinate directly: `bcoord = out_tidx / blockSize` where `out_tidx` is the 4D tensor coordinate in WHCN layout. The linear block ID calculation uses the same pre-computed stride approach: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`. The element value is loaded using the corresponding input buffer index: `value = t_in[in_bufi]` where `in_bufi = tidx_to_bufi(out_tidx, t_in_strides)`. Quantization applies the block-specific parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])`. **Future Improvements**: Dynamic workgroup sizing based on block dimensions, there is likely a better method to making it better than what it is currently. ghstack-source-id: 299473617 @exported-using-ghexport Differential Revision: [D78302195](https://our.internmc.facebook.com/intern/diff/D78302195/)
…anup Pull Request resolved: #12576 # Changes * Implement `torchao.dequantize_affine` operator in Vulkan backend with comprehensive texture and buffer storage support * Add block-wise dequantization mode in `dequantize_texture.glsl` and `dequantize_buffer.glsl` shaders for configurable tensor block dequantization * Extend dequantization infrastructure in `Dequantize.cpp` to handle affine transformations with configurable block sizes and quantization parameters * Support integer-to-floating-point conversion with precise reconstruction of original values BE: Improved the documentation in the shader logic which is more detailed and clear # Motivation The existing Vulkan quantization infrastructure lacked support for the `torchao.dequantize_affine` operator, which is essential for completing the quantization-dequantization cycle in dynamic quantization workflows. The `dequantize_affine` operator provides flexible block-wise dequantization that reconstructs floating-point values from quantized integer blocks, enabling: * **Block-wise Dequantization**: Reconstructs floating-point values from configurable tensor blocks using separate scale and zero-point parameters, enabling precise recovery of original data distributions * **Affine Transformation**: Uses the formula `value = (qvalue - zero_point) * scale` for accurate integer-to-floating-point mapping # Operator Description The `dequantize_affine` operator converts n-bit integer tensor values back to floating-point representations using pre-computed quantization parameters (scale and zero_point) applied to configurable tensor blocks. Block-wise dequantization divides tensors into blocks and applies separate dequantization parameters to each block, allowing fine-grained reconstruction of the original floating-point precision. The dequantization formula is: `value = (qvalue - zero_point) * scale` **Storage Requirements**: Scale and zero_point tensors must use buffer storage with width-packed layout. Input/output tensors support both buffer and texture storage with standard axis mapping. Input tensors must be integer types (kByte, kChar, kInt). # Block-wise Dequantization Implementation Block-wise dequantization enables fine-grained reconstruction by dividing tensors into blocks and applying separate dequantization parameters to each block. The implementation uses the same key data structures computed in `Dequantize.cpp`: * **`block_size_vec`**: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks) * **`tensor_size_whcn`**: Input tensor dimensions converted to WHCN layout using `utils::make_whcn_ivec4()` * **`num_blocks_vec`**: Number of blocks per dimension calculated as `tensor_size_whcn / block_size_vec` * **`block_stride_vec`**: Pre-computed linear strides for block grid indexing `{1, #W, #W*#H, #W*#H*#C}` to enable efficient block ID calculation The block coordinate calculation uses: `bcoord = tidx / blockSize` where `tidx` is the tensor coordinate in WHCN layout, then the linear block ID is computed as: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w` # Shader Algorithm Overview ## Texture Storage Implementation (`dequantize_texture.glsl`) **Workgroup Configuration**: - **Global WG Size**: Default sizing based on texture dimensions - **Local WG Size**: Default with special handling for batch dimension dequantization (Z dimension set to 1 for proper workgroup dispatching when `global_workgroup_size[2] > 1`) **Block-wise Mode Algorithm**: The shader processes 3D texture positions where each position represents a texel containing 4 width-packed integer components. For each texel at position `pos`, it calculates a base tensor index `base_tidx = ivec4(pos.x * 4, pos.y, pos.z, 0)` to account for width-packing. For each of the 4 components in the texel, it computes the actual tensor coordinate: `tidx = ivec4(base_tidx.x + i, base_tidx.y, (foldedZ % C_total), (foldedZ / C_total))` where `foldedZ = pos.z` handles batch-channel folding in 4D tensors and `C_total = numBlocks.z * blockSize.z` represents the total channel dimension. The block coordinate is calculated using integer division: `bcoord = tidx / blockSize`, then the linear block ID uses pre-computed strides: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`. Each integer component is dequantized using its corresponding block's parameters: `value = dequantize_val(qvalue, t_scale[block_id], t_zero_point[block_id])` where `dequantize_val()` applies the formula `(qvalue - zero_point) * scale`. The reconstructed floating-point values are written to the output texel with proper type handling for double precision outputs. ## Buffer Storage Implementation (`dequantize_buffer.glsl`) **Workgroup Configuration**: - **Global WG Size**: Default sizing based on buffer element count - **Local WG Size**: Default sizing without special constraints **Block-wise Mode Algorithm**: The shader processes linear buffer indices using `gl_GlobalInvocationID.x` as the output buffer index. It converts this to tensor coordinates using `bufi_to_tidx(out_bufi, t_out_strides, out_dim_order)` which handles the buffer-to-tensor index mapping with proper stride calculations. For each element, it computes the block coordinate directly: `bcoord = out_tidx / blockSize` where `out_tidx` is the 4D tensor coordinate in WHCN layout. The linear block ID calculation uses the same pre-computed stride approach: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`. The quantized integer value is loaded using the corresponding input buffer index: `qvalue = t_in[in_bufi]` where `in_bufi = tidx_to_bufi(out_tidx, t_in_strides)`. Dequantization applies the block-specific parameters: `value = dequantize_val(qvalue, t_scale[block_id], t_zero_point[block_id])` to reconstruct the original floating-point value. **Future Improvements**: Dynamic workgroup sizing based on block dimensions ghstack-source-id: 299473614 @exported-using-ghexport Differential Revision: [D78435552](https://our.internmc.facebook.com/intern/diff/D78435552/)
Pull Request resolved: #12577 # Changes * Implement `torchao.choose_qparams_affine` operator in Vulkan backend with comprehensive buffer storage support * Add block-wise quantization parameter computation in `choose_qparams_buffer.glsl` shader for configurable tensor block analysis * Extend quantization parameter infrastructure in `ChooseQParams.cpp` to handle affine transformations with configurable block sizes and multiple mapping types * Support three quantization mapping strategies: ASYMMETRIC, SYMMETRIC, and SYMMETRIC_NO_CLIPPING_ERR for optimal parameter selection * Consolidated the logic for choosing scale and zero point between affine cases and regular quantized_decomposed cases. BE: Improved the documentation in the shader logic which is more detailed and clear # Motivation The existing Vulkan quantization infrastructure lacked support for the `torchao.choose_qparams_affine` operator, which is essential for computing optimal quantization parameters in dynamic quantization workflows. The `choose_qparams_affine` operator provides flexible block-wise parameter computation that analyzes statistical distributions within tensor blocks, enabling: * **Block-wise Parameter Computation**: Analyzes configurable tensor blocks to compute optimal scale and zero-point values, improving quantization accuracy for heterogeneous data distributions * **Multiple Mapping Types**: Supports ASYMMETRIC, SYMMETRIC, and SYMMETRIC_NO_CLIPPING_ERR quantization strategies for different precision-performance trade-offs # Operator Description The `choose_qparams_affine` operator computes optimal quantization parameters (scale and zero_point) from floating-point tensor blocks using statistical analysis of data distributions. Block-wise parameter computation divides tensors into blocks and analyzes each block independently to determine the best quantization mapping for subsequent quantization operations. The parameter calculation varies by mapping type: - **ASYMMETRIC**: `scale = (max - min) / (quant_max - quant_min)`, `zero_point = quant_min - round(min / scale)` - **SYMMETRIC**: `scale = max_abs / ((quant_max - quant_min) / 2)`, `zero_point = midpoint` - **SYMMETRIC_NO_CLIPPING_ERR**: `scale = max(abs(min)/abs(quant_min), max/quant_max)`, `zero_point = midpoint` **Storage Requirements**: Input tensors must be floating-point (kFloat) with width-packed layout. Output scale/zero_point tensors use buffer storage. NOTE: Texture storage implementation is not supported due to complexity of block-wise coordinate mapping in 3D texture space. This will likely be necessary for better efficiency in the future. # Block-wise Parameter Computation Implementation Block-wise parameter computation enables fine-grained quantization analysis by dividing tensors into blocks and computing separate scale/zero_point parameters for each block. The implementation uses several key data structures computed in `ChooseQParams.cpp`: * **`block_size_vec`**: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks) * **`tensor_size_whcn`**: Input tensor dimensions converted to WHCN layout using `utils::make_whcn_ivec4()` * **`num_blocks_vec`**: Number of blocks per dimension calculated as `ceil(tensor_size_whcn / block_size_vec)` to handle non-divisible dimensions * **`block_stride_vec`**: Pre-computed linear strides for block grid indexing `{1, #W, #W*#H, #W*#H*#C}` to enable efficient block ID calculation * **`mapping_type`**: Integer encoding of quantization strategy (0=ASYMMETRIC, 1=SYMMETRIC, 2=SYMMETRIC_NO_CLIPPING_ERR) The block coordinate calculation uses: `block_coord = block_id_to_coord(block_id)` which converts linear block IDs back to 4D WHCN coordinates, then computes element ranges: `t0 = block_coord * blockSize` and `tEnd = t0 + blockSize` for nested loop iteration. # Shader Algorithm Overview ## Buffer Storage Implementation (`choose_qparams_buffer.glsl`) **Workgroup Configuration**: - **Global WG Size**: `{nBlocks, 1u, 1u}` where `nBlocks = total number of blocks` computed from `ceil(tensor_size / block_size)` for each dimension - **Local WG Size**: `{1u, 1u, 1u}` (single thread per block for simplicity, though could be optimized for larger blocks) **Block-wise Mode Algorithm**: The shader uses a sophisticated multi-level nested approach to process tensor blocks efficiently. Each thread is assigned multiple blocks using strided access: `for (uint block_id = gl_GlobalInvocationID.x; block_id < TOTAL_BLOCKS; block_id += STRIDE)` where `STRIDE = gl_WorkGroupSize.x * gl_NumWorkGroups.x`. For each assigned block, the algorithm performs several key steps: **1. Block Coordinate Conversion**: The `block_id_to_coord(block_id)` function converts linear block IDs to 4D WHCN coordinates using modular arithmetic. **2. Element Range Calculation**: Computes the inclusive start coordinate `t0 = bc * blockSize` and exclusive end coordinate `tEnd = t0 + blockSize` to define the block's element boundaries in tensor space. **3. Nested Loop Min/Max Scan**: Uses four nested loops to iterate through all elements within the block: `for (int n = t0.w; n < tEnd.w; ++n) for (int c = t0.z; c < tEnd.z; ++c) for (int h = t0.y; h < tEnd.y; ++h) for (int w = t0.x; w < tEnd.x; ++w)` Each element is accessed using `tidx_to_bufi(ivec4(w,h,c,n), t_in_strides)` to convert 4D tensor coordinates to linear buffer indices with proper stride handling. **4. Parameter Calculation**: Calls `calc_scale_zp(lo, hi, quant_min, quant_max, mapping_type, eps, scale, zp)` which implements the three mapping strategies: * **ASYMMETRIC (mapping_type=0)**: Maps full range [min, max] to [quant_min, quant_max] preserving data distribution * **SYMMETRIC (mapping_type=1)**: Centers around zero using `max_abs = max(abs(min), abs(max))` for balanced quantization * **SYMMETRIC_NO_CLIPPING_ERR (mapping_type=2)**: Computes separate scales for positive/negative ranges and uses the maximum to prevent clipping **Future Improvements**: Implement workgroup-level reduction for large blocks, optimize memory access patterns for better cache utilization, and explore texture storage implementation with simplified block alignment constraints. ghstack-source-id: 299473615 @exported-using-ghexport Differential Revision: [D78436638](https://our.internmc.facebook.com/intern/diff/D78436638/)
pytorch-bot
bot
added
the
module: vulkan
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/13003
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #12577 by @ahmtox
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/ahmtox/45/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/ahmtox/45/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/ahmtox/44/orig
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/ahmtox/45/orig
@diff-train-skip-merge
cc @SS-JIA @manuelcandales @cbilgin