hip-quant 2.2.1__tar.gz → 2.2.2__tar.gz

{hip_quant-2.2.1/hip_quant.egg-info → hip_quant-2.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hip-quant
-Version: 2.2.1
+Version: 2.2.2
 Summary: High-performance GPU tensor quantization and inference engine for AMD ROCm. Native AOT GEMV (K-Quants & I-Quants), State Space Model (SSM/DeltaNet) kernels, FP8 WMMA WaveAttention, and portable HIP C++ kernels for RDNA4, RDNA3, and CDNA.
 Author-email: Thenukegun10x <conormorris84@gmail.com>
 License:                                  Apache License
@@ -239,6 +239,14 @@ Dynamic: license-file
 
 `hip-quant` is a standalone Python library and highly optimized HIP C++ backend for tensor quantization and accelerated LLM inference on AMD GPUs with zero CPU round-trips.
 
+## 🌟 What's New in 2.2.2
+
+> **2.2.2.post215** is the same release rebuilt against PyTorch 2.15 (TheRock / ROCm 7.14).
+
+- **Fixed a packaging bug that silently disabled the gfx12 FP8 WMMA path.** `codebooks/iq2_s.bin` was matched by `.gitignore` and never tracked, so it was missing from git, the sdist, and every wheel built from a clean tree. `ensure_initialized()` requires all six I-Quant codebooks and returned failure, which made `get_arch_name`/`get_device_name`/`get_selected_device` report empty — and the `gfx12` gate then failed closed. The codebook is now tracked and shipped, and `codebooks/` is no longer ignored.
+- **Device detection is decoupled from codebook loading.** `get_device_name`, `get_arch_name`, `get_selected_device`, `get_device_memory` and `device_has_wmma` now only require the device to initialise, so a missing or corrupt codebook fails the I-Quant formats loudly instead of bricking GPU detection for every format.
+- The `_C` extension is unchanged from 2.2.1; the Windows `hip_quantize*.dll` were rebuilt with the init split (both all-arch targets).
+
 ## 🌟 What's New in 2.2.1
 
 > **2.2.1.post215** is the same release rebuilt against PyTorch 2.15 (TheRock / ROCm 7.14). See *Custom PyTorch 2.15* below for a self-hosted install index.
@@ -574,11 +582,11 @@ matches your PyTorch, so the extension loads without an ABI mismatch:
 
 | Wheel | PyTorch | ROCm |
 |---|---|---|
-| `hip-quant==2.2.1` | 2.9.x | ROCm 7.2.1 |
-| `hip-quant==2.2.1.post215` | 2.15 / TheRock | ROCm 7.14 |
+| `hip-quant==2.2.2` | 2.9.x | ROCm 7.2.1 |
+| `hip-quant==2.2.2.post215` | 2.15 / TheRock | ROCm 7.14 |
 
 ```powershell
-pip install "hip-quant==2.2.1.post215"   # PyTorch 2.15 / TheRock
+pip install "hip-quant==2.2.2.post215"   # PyTorch 2.15 / TheRock
 ```
 
 #### Custom PyTorch 2.15 (self-hosted wheels)
@@ -588,7 +596,7 @@ Its wheels live in a public Backblaze B2 bucket; point pip at the index page
 with `--find-links`:
 
 ```powershell
-pip install "hip-quant==2.2.1.post215" `
+pip install "hip-quant==2.2.2.post215" `
   --find-links https://dl.hipquant.download/file/Torchs/torch_index.html
 ```
 

{hip_quant-2.2.1 → hip_quant-2.2.2}/README.md

@@ -14,6 +14,14 @@
 
 `hip-quant` is a standalone Python library and highly optimized HIP C++ backend for tensor quantization and accelerated LLM inference on AMD GPUs with zero CPU round-trips.
 
+## 🌟 What's New in 2.2.2
+
+> **2.2.2.post215** is the same release rebuilt against PyTorch 2.15 (TheRock / ROCm 7.14).
+
+- **Fixed a packaging bug that silently disabled the gfx12 FP8 WMMA path.** `codebooks/iq2_s.bin` was matched by `.gitignore` and never tracked, so it was missing from git, the sdist, and every wheel built from a clean tree. `ensure_initialized()` requires all six I-Quant codebooks and returned failure, which made `get_arch_name`/`get_device_name`/`get_selected_device` report empty — and the `gfx12` gate then failed closed. The codebook is now tracked and shipped, and `codebooks/` is no longer ignored.
+- **Device detection is decoupled from codebook loading.** `get_device_name`, `get_arch_name`, `get_selected_device`, `get_device_memory` and `device_has_wmma` now only require the device to initialise, so a missing or corrupt codebook fails the I-Quant formats loudly instead of bricking GPU detection for every format.
+- The `_C` extension is unchanged from 2.2.1; the Windows `hip_quantize*.dll` were rebuilt with the init split (both all-arch targets).
+
 ## 🌟 What's New in 2.2.1
 
 > **2.2.1.post215** is the same release rebuilt against PyTorch 2.15 (TheRock / ROCm 7.14). See *Custom PyTorch 2.15* below for a self-hosted install index.
@@ -349,11 +357,11 @@ matches your PyTorch, so the extension loads without an ABI mismatch:
 
 | Wheel | PyTorch | ROCm |
 |---|---|---|
-| `hip-quant==2.2.1` | 2.9.x | ROCm 7.2.1 |
-| `hip-quant==2.2.1.post215` | 2.15 / TheRock | ROCm 7.14 |
+| `hip-quant==2.2.2` | 2.9.x | ROCm 7.2.1 |
+| `hip-quant==2.2.2.post215` | 2.15 / TheRock | ROCm 7.14 |
 
 ```powershell
-pip install "hip-quant==2.2.1.post215"   # PyTorch 2.15 / TheRock
+pip install "hip-quant==2.2.2.post215"   # PyTorch 2.15 / TheRock
 ```
 
 #### Custom PyTorch 2.15 (self-hosted wheels)
@@ -363,7 +371,7 @@ Its wheels live in a public Backblaze B2 bucket; point pip at the index page
 with `--find-links`:
 
 ```powershell
-pip install "hip-quant==2.2.1.post215" `
+pip install "hip-quant==2.2.2.post215" `
   --find-links https://dl.hipquant.download/file/Torchs/torch_index.html
 ```
 

{hip_quant-2.2.1 → hip_quant-2.2.2}/__init__.py

@@ -3,7 +3,7 @@ import numpy as np
 import os
 import sys
 
-__version__ = "2.2.1"
+__version__ = "2.2.2"
 
 
 _TORCH_EXPORTS = {

{hip_quant-2.2.1 → hip_quant-2.2.2/hip_quant.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hip-quant
-Version: 2.2.1
+Version: 2.2.2
 Summary: High-performance GPU tensor quantization and inference engine for AMD ROCm. Native AOT GEMV (K-Quants & I-Quants), State Space Model (SSM/DeltaNet) kernels, FP8 WMMA WaveAttention, and portable HIP C++ kernels for RDNA4, RDNA3, and CDNA.
 Author-email: Thenukegun10x <conormorris84@gmail.com>
 License:                                  Apache License
@@ -239,6 +239,14 @@ Dynamic: license-file
 
 `hip-quant` is a standalone Python library and highly optimized HIP C++ backend for tensor quantization and accelerated LLM inference on AMD GPUs with zero CPU round-trips.
 
+## 🌟 What's New in 2.2.2
+
+> **2.2.2.post215** is the same release rebuilt against PyTorch 2.15 (TheRock / ROCm 7.14).
+
+- **Fixed a packaging bug that silently disabled the gfx12 FP8 WMMA path.** `codebooks/iq2_s.bin` was matched by `.gitignore` and never tracked, so it was missing from git, the sdist, and every wheel built from a clean tree. `ensure_initialized()` requires all six I-Quant codebooks and returned failure, which made `get_arch_name`/`get_device_name`/`get_selected_device` report empty — and the `gfx12` gate then failed closed. The codebook is now tracked and shipped, and `codebooks/` is no longer ignored.
+- **Device detection is decoupled from codebook loading.** `get_device_name`, `get_arch_name`, `get_selected_device`, `get_device_memory` and `device_has_wmma` now only require the device to initialise, so a missing or corrupt codebook fails the I-Quant formats loudly instead of bricking GPU detection for every format.
+- The `_C` extension is unchanged from 2.2.1; the Windows `hip_quantize*.dll` were rebuilt with the init split (both all-arch targets).
+
 ## 🌟 What's New in 2.2.1
 
 > **2.2.1.post215** is the same release rebuilt against PyTorch 2.15 (TheRock / ROCm 7.14). See *Custom PyTorch 2.15* below for a self-hosted install index.
@@ -574,11 +582,11 @@ matches your PyTorch, so the extension loads without an ABI mismatch:
 
 | Wheel | PyTorch | ROCm |
 |---|---|---|
-| `hip-quant==2.2.1` | 2.9.x | ROCm 7.2.1 |
-| `hip-quant==2.2.1.post215` | 2.15 / TheRock | ROCm 7.14 |
+| `hip-quant==2.2.2` | 2.9.x | ROCm 7.2.1 |
+| `hip-quant==2.2.2.post215` | 2.15 / TheRock | ROCm 7.14 |
 
 ```powershell
-pip install "hip-quant==2.2.1.post215"   # PyTorch 2.15 / TheRock
+pip install "hip-quant==2.2.2.post215"   # PyTorch 2.15 / TheRock
 ```
 
 #### Custom PyTorch 2.15 (self-hosted wheels)
@@ -588,7 +596,7 @@ Its wheels live in a public Backblaze B2 bucket; point pip at the index page
 with `--find-links`:
 
 ```powershell
-pip install "hip-quant==2.2.1.post215" `
+pip install "hip-quant==2.2.2.post215" `
   --find-links https://dl.hipquant.download/file/Torchs/torch_index.html
 ```
 

{hip_quant-2.2.1 → hip_quant-2.2.2}/hip_quant.egg-info/SOURCES.txt

@@ -56,6 +56,7 @@ wave_helpers.py
 ./torch_api.py
 ./wave_helpers.py
 ./codebooks/iq1_s.bin
+./codebooks/iq2_s.bin
 ./codebooks/iq2_xs.bin
 ./codebooks/iq2_xxs.bin
 ./codebooks/iq3_s.bin
@@ -171,6 +172,7 @@ wave_helpers.py
 ./torch_ext/wave_attn_long.hip
 ./torch_ext/wave_attn_prefill.hip
 codebooks/iq1_s.bin
+codebooks/iq2_s.bin
 codebooks/iq2_xs.bin
 codebooks/iq2_xxs.bin
 codebooks/iq3_s.bin

{hip_quant-2.2.1 → hip_quant-2.2.2}/hip_quantize.cpp

@@ -259,7 +259,10 @@ static bool load_iq_host_codebook(
     return true;
 }
 
-static bool ensure_initialized() {
+// Device-only init: enumerate, select, and cache device properties. Kept
+// separate from I-Quant codebook loading so a missing codebook cannot brick
+// device detection (arch/name/selected_device) for every other format.
+static bool ensure_device() {
     if (!hip_initialized) {
         int count = 0;
         if (!hip_check(hipGetDeviceCount(&count), "hipGetDeviceCount")) return false;
@@ -297,6 +300,14 @@ static bool ensure_initialized() {
         if (!hip_check(hipGetDeviceProperties(&props, device_id), "hipGetDeviceProperties")) return false;
         hip_initialized = true;
     }
+    return true;
+}
+
+// Full init: device + all I-Quant codebooks. Used by the format entry points
+// that may need a codebook, so a missing file fails those calls loudly instead
+// of letting kernels read un-uploaded device tables and return wrong results.
+static bool ensure_initialized() {
+    if (!ensure_device()) return false;
     if (!iq3xxs_tables_loaded) {
         iq_host_codebook table = {};
         if (!load_iq_host_codebook("iq3_xxs", 256 * 4, 4096 * sizeof(int), 22825 * sizeof(uint16_t), &table)) return false;
@@ -544,7 +555,7 @@ HIP_QUANT_EXPORT size_t ggml_row_size_for(int type, int64_t n_per_row) {
 
 HIP_QUANT_EXPORT const char* get_device_name() {
     static thread_local char name[256];
-    if (!ensure_initialized()) return "";
+    if (!ensure_device()) return "";
     strncpy(name, props.name, 255);
     name[255] = '\0';
     return name;
@@ -1839,7 +1850,7 @@ HIP_QUANT_EXPORT int get_device_prop(
 }
 
 HIP_QUANT_EXPORT int get_arch_name(char *buf, int buf_size) {
-    if (!ensure_initialized()) return 1;
+    if (!ensure_device()) return 1;
     strncpy(buf, props.gcnArchName, (size_t)(buf_size - 1));
     buf[buf_size - 1] = '\0';
     return 0;
@@ -1849,7 +1860,7 @@ HIP_QUANT_EXPORT int get_arch_name(char *buf, int buf_size) {
 // index is relative to HIP_VISIBLE_DEVICES when that environment variable is
 // set before the process starts.
 HIP_QUANT_EXPORT int get_selected_device() {
-    return ensure_initialized() ? device_id : -1;
+    return ensure_device() ? device_id : -1;
 }
 
 HIP_QUANT_EXPORT int get_hip_runtime_version() {
@@ -1859,7 +1870,7 @@ HIP_QUANT_EXPORT int get_hip_runtime_version() {
 }
 
 HIP_QUANT_EXPORT int get_device_memory(size_t *free_bytes, size_t *total_bytes) {
-    if (!ensure_initialized()) {
+    if (!ensure_device()) {
         *free_bytes = 0;
         *total_bytes = 0;
         return 1;
@@ -1874,7 +1885,7 @@ HIP_QUANT_EXPORT int get_device_memory(size_t *free_bytes, size_t *total_bytes)
 }
 
 HIP_QUANT_EXPORT int device_has_wmma() {
-    if (!ensure_initialized()) return 0;
+    if (!ensure_device()) return 0;
     const char *arch = props.gcnArchName;
     // This reports support for the FP8/BF8 gfx12 w32 intrinsics used by
     // fp8_gemm_wmma_kernel, not general matrix-core or rocWMMA capability.

{hip_quant-2.2.1 → hip_quant-2.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hip-quant"
-version = "2.2.1"
+version = "2.2.2"
 description = "High-performance GPU tensor quantization and inference engine for AMD ROCm. Native AOT GEMV (K-Quants & I-Quants), State Space Model (SSM/DeltaNet) kernels, FP8 WMMA WaveAttention, and portable HIP C++ kernels for RDNA4, RDNA3, and CDNA."
 readme = "README.md"