Skip to content

Commit 62e3409

Browse files
ElectroiscodingElectroiscoding
committed
cuda yes
1 parent 4479a67 commit 62e3409

File tree

7 files changed

+816
-116
lines changed

7 files changed

+816
-116
lines changed

‎Crayon_Colab_Notebook.py‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,5 @@
11
"""
2-
XERV CRAYON V4.3.1 - Production Omni-Backend Tokenizer
2+
XERV CRAYON V4.1.9 - Production Omni-Backend Tokenizer
33
=======================================================
44
Copy this ENTIRE script into a Google Colab cell and run it.
55
@@ -13,7 +13,7 @@
1313
import time
1414

1515
print("=" * 70)
16-
print("XERV CRAYON V4.3.1 INSTALLATION")
16+
print("XERV CRAYON V4.1.9 INSTALLATION AND BENCHMARKS")
1717
print("=" * 70)
1818

1919
# 1. Environment Check

‎README.md‎

Lines changed: 80 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,7 @@
22
<img src="https://em-content.zobj.net/source/microsoft-teams/363/crayon_1f58d-fe0f.png" width="120" alt="Crayon Logo"/>
33
</p>
44

5-
<h1 align="center">🖍️ XERV Crayon v4.0</h1>
5+
<h1 align="center">🖍️ XERV Crayon v4.1.9</h1>
66

77
<p align="center">
88
<strong>The Omni-Backend Tokenizer for Specialized AI</strong>
@@ -30,40 +30,98 @@
3030
|:--------|:------------|
3131
| **💾 Cartridge System** | Instantly hot-swap specialized vocabularies (`science`, `code`, `multilingual`) |
3232
| **🚀 Omni-Backend** | Auto-detects & runs on **CPU (AVX2)**, **NVIDIA (CUDA)**, or **AMD (ROCm)** |
33-
| **⚡ Native GPU Kernels** | "Bare Metal" C++/HIP kernels (no wrappers) for >100M tokens/sec |
33+
| **⚡ Native GPU Kernels** | "Bare Metal" C++/CUDA/HIP kernels (no wrappers) for >10M tokens/sec |
3434
| **🗺️ Zero-Copy Mapping** | DAT files loaded via `mmap` for instant startup & minimal RAM |
3535
| **🌊 Zero-Disk Streaming** | Build profiles directly from Hugging Face—no multi-GB downloads |
3636
| **🛡️ Offline Resilience** | Seamless local bootstrap fallback. Works offline out-of-the-box |
3737

3838
---
3939

40-
## 📊 Benchmarks — The Numbers Speak
40+
## 📊 Benchmarks — Production Results (Tesla T4 GPU)
4141

42-
> **100% HONEST. NO SUGARCOATING. DATA-DRIVEN.**
42+
> **100% VERIFIED. GOOGLE COLAB T4 GPU.**
4343
>
44-
> Run `python benchmark_competitive.py` to reproduce these results yourself.
44+
> Complete installation and benchmark logs from actual T4 GPU testing.
4545
46-
### ⚡ Speed Comparison (Omni-Backend)
46+
### ⚡ Installation Summary (T4 GPU Environment)
4747

48-
| Tokenizer | Tokens/sec | vs CRAYON |
49-
|:----------|----------:|:----------|
50-
| **🖍️ CRAYON (CPU - AVX2)** | **21,863,777** | **baseline** |
51-
| **🖍️ CRAYON (CUDA - A100)** | **140,000,000+** | **6.4x faster** |
52-
| tiktoken (GPT-4) | 524,469 | 41x slower |
53-
| HF LLaMA (SP-BPE) | 281,558 | 77x slower |
54-
| HF GPT-2 (BPE) | 237,117 | 92x slower |
55-
| HF BERT (WordPiece) | 202,269 | 108x slower |
48+
```
49+
======================================================================
50+
XERV CRAYON V4.1.9 INSTALLATION AND BENCHMARKS
51+
======================================================================
52+
[1/7] Checking environment...
53+
PyTorch: 2.9.0+cu126
54+
CUDA: 12.6 (Tesla T4)
55+
* Smart Build: Will compile ONLY for this GPU architecture
56+
NVCC: /usr/local/cuda/bin/nvcc
57+
58+
[2/7] Installing build dependencies...
59+
Done (ninja, packaging, wheel)
60+
61+
[3/7] Cleaning previous installations...
62+
63+
[4/7] Cloning source code...
64+
__version__ = "4.1.9"
65+
66+
[5/7] Compiling and Installing (Streaming Logs)...
67+
----------------------------------------------------------------------
68+
[CRAYON-BUILD] Detected GPU: SM 7.5 -> Compiling for sm_75 ONLY
69+
[CRAYON-BUILD] Configuring CUDA extension (max_jobs=1)
70+
71+
building 'crayon.c_ext.crayon_cpu' extension
72+
[1/1] c++ -O3 -march=native -mavx2 -fPIC -std=c++17
73+
Successfully built crayon_cpu.so
74+
75+
building 'crayon.c_ext.crayon_cuda' extension
76+
[1/1] nvcc -O3 -std=c++17 --expt-relaxed-constexpr -gencode=arch=compute_75,code=sm_75
77+
Successfully built crayon_cuda.so
78+
79+
Successfully installed xerv-crayon-4.1.9
80+
----------------------------------------------------------------------
81+
82+
[6/7] Verifying installation...
83+
Success! Installed version: 4.1.9
84+
Backends: {'cpu': True, 'cuda': True, 'rocm': False}
85+
```
86+
87+
### 🔥 Performance Results (T4 GPU vs Tiktoken)
88+
89+
**CRAYON (CUDA Backend - Tesla T4):**
90+
```
91+
Active Device: CUDA
92+
Backend: cuda_extension
93+
94+
Batch Throughput (XERV CRAYON):
95+
1,000 docs: 748,048 docs/sec | 9,724,621 tokens/sec
96+
10,000 docs: 639,239 docs/sec | 8,310,109 tokens/sec
97+
50,000 docs: 781,129 docs/sec | 10,154,678 tokens/sec
98+
```
99+
100+
**Tiktoken (cl100k_base - CPU):**
101+
```
102+
Tiktoken Batch Throughput (cl100k_base encoding):
103+
1,000 docs: 87,307 docs/sec | 873,068 tokens/sec
104+
10,000 docs: 81,658 docs/sec | 816,576 tokens/sec
105+
50,000 docs: 107,583 docs/sec | 1,075,829 tokens/sec
106+
```
107+
108+
### 📈 Performance Comparison Table
109+
110+
| Batch Size | CRAYON Docs/Sec | CRAYON Tokens/Sec | Tiktoken Docs/Sec | Tiktoken Tokens/Sec | **Speedup** |
111+
|:-----------|----------------:|------------------:|------------------:|--------------------:|------------:|
112+
| 1,000 | 748,048 | 9,724,621 | 87,307 | 873,068 | **11.1x** ✨ |
113+
| 10,000 | 639,239 | 8,310,109 | 81,658 | 816,576 | **10.2x** ✨ |
114+
| 50,000 | 781,129 | 10,154,678 | 107,583 | 1,075,829 | **9.4x** ✨ |
56115

57-
### 📈 CPU Optimization Verification
58-
*Measured on Intel Core i3-7020U (Low-Power Laptop CPU)*
116+
**Average Speedup: 10.2x faster than tiktoken on Tesla T4 GPU**
59117

60-
| Metric | Result |
61-
|:-------|:-------|
62-
| ✅ **AVX2 Status** | Active (Simd-Ops v4) |
63-
| ✅ **Load Time** | **0.54ms** (Instant hot-swap) |
64-
| ✅ **Throughput** | **21.1M tokens/sec** (!?!) |
118+
### 🎯 Key Achievements
65119

66-
![Benchmark Comparison](benchmark_comparison.png)
120+
- ✅ **>10M tokens/sec** on mid-tier GPU (Tesla T4)
121+
- ✅ **Smart compilation** - Only builds for detected GPU architecture
122+
- ✅ **Zero-copy memory mapping** - Instant profile loading (<1ms)
123+
- ✅ **Production-grade stability** - Handles 50K+ document batches
124+
- ✅ **Consistent performance** - Minimal variance across batch sizes
67125

68126
---
69127

‎RELEASE_NOTES_4.1.9.md‎

Lines changed: 194 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,194 @@
1+
# XERV CRAYON V4.1.9 - Release Summary
2+
3+
## 🎉 Successfully Published to PyPI!
4+
5+
**Package URL:** https://pypi.org/project/xerv-crayon/4.1.9/
6+
7+
---
8+
9+
## 📦 Installation
10+
11+
```bash
12+
pip install xerv-crayon
13+
```
14+
15+
For Google Colab with GPU:
16+
```python
17+
# Copy and run Crayon_Colab_Notebook.py or colab_benchmark.py
18+
```
19+
20+
---
21+
22+
## 🚀 Local Benchmark Results (Your Machine)
23+
24+
### Hardware Configuration
25+
- **OS:** Windows 10.0.19045
26+
- **Python:** 3.13.1
27+
- **CPU:** Intel (AVX2 enabled)
28+
- **GPU:** Not available (CPU-only benchmarks)
29+
30+
### Performance Results
31+
32+
**CRAYON (CPU Backend - AVX2):**
33+
```
34+
Batch Throughput (CPU):
35+
1,000 docs: 842,230 docs/sec | 10,948,986 tokens/sec
36+
10,000 docs: 560,384 docs/sec | 7,284,988 tokens/sec
37+
50,000 docs: 447,427 docs/sec | 5,816,548 tokens/sec
38+
```
39+
40+
**Tiktoken (cl100k_base - CPU):**
41+
```
42+
Tiktoken Batch Throughput:
43+
1,000 docs: 11,007 docs/sec | 110,069 tokens/sec
44+
10,000 docs: 12,861 docs/sec | 128,610 tokens/sec
45+
50,000 docs: 13,386 docs/sec | 133,865 tokens/sec
46+
```
47+
48+
### Performance Summary
49+
50+
| Batch Size | CRAYON Tokens/Sec | Tiktoken Tokens/Sec | **Speedup** |
51+
|:-----------|------------------:|--------------------:|------------:|
52+
| 1,000 | 10,948,986 | 110,069 | **99.5x** ✨ |
53+
| 10,000 | 7,284,988 | 128,610 | **56.6x** ✨ |
54+
| 50,000 | 5,816,548 | 133,865 | **43.5x** ✨ |
55+
56+
**Average Speedup: 64.6x faster than tiktoken on CPU**
57+
58+
---
59+
60+
## 🔥 Google Colab T4 GPU Results (Included in README)
61+
62+
**CRAYON (CUDA Backend - Tesla T4):**
63+
```
64+
Batch Throughput:
65+
1,000 docs: 748,048 docs/sec | 9,724,621 tokens/sec
66+
10,000 docs: 639,239 docs/sec | 8,310,109 tokens/sec
67+
50,000 docs: 781,129 docs/sec | 10,154,678 tokens/sec
68+
```
69+
70+
**Average Speedup: 10.2x faster than tiktoken on T4 GPU**
71+
72+
---
73+
74+
## 📝 Files Updated
75+
76+
### Version Updates
77+
- ✅ `src/crayon/__init__.py` - Updated to v4.1.9
78+
- ✅ `pyproject.toml` - Updated to v4.1.9
79+
80+
### New Files Created
81+
- ✅ `local_benchmark.py` - Comprehensive local benchmarking with hardware detection
82+
- ✅ `colab_benchmark.py` - Production-grade Colab installation and benchmark script
83+
- ✅ `Crayon_Colab_Notebook.py` - Updated to v4.1.9
84+
85+
### Documentation Updates
86+
- ✅ `README.md` - Complete rewrite of hero section with T4 GPU benchmark results
87+
- Added detailed installation logs
88+
- Added performance comparison tables
89+
- Added key achievements section
90+
- Removed old benchmark data
91+
- Added production-verified results
92+
93+
---
94+
95+
## 🎯 Key Features of This Release
96+
97+
1. **Production-Grade Benchmarking**
98+
- Deep hardware detection (CPU model, cores, frequency, GPU info)
99+
- Windows/Linux compatible
100+
- ASCII-safe output (no Unicode issues)
101+
- Automatic backend detection
102+
103+
2. **Comprehensive Testing**
104+
- Local CPU benchmarks
105+
- Google Colab GPU benchmarks
106+
- Tiktoken comparison
107+
- Multiple batch sizes (1K, 10K, 50K documents)
108+
109+
3. **Clean, Readable Code**
110+
- Minimal comments
111+
- Clear function names
112+
- Production-grade error handling
113+
- No placeholders or pseudocode
114+
115+
4. **PyPI Publishing**
116+
- Successfully published to PyPI
117+
- Version 4.1.9
118+
- Includes both source distribution and wheel
119+
120+
---
121+
122+
## 🔧 Usage Examples
123+
124+
### Quick Start
125+
```python
126+
from crayon import CrayonVocab
127+
128+
vocab = CrayonVocab(device="auto")
129+
vocab.load_profile("lite")
130+
131+
text = "Hello, world!"
132+
tokens = vocab.tokenize(text)
133+
print(tokens)
134+
```
135+
136+
### Batch Processing
137+
```python
138+
from crayon import CrayonVocab
139+
140+
vocab = CrayonVocab(device="cpu")
141+
vocab.load_profile("code")
142+
143+
documents = ["def hello():", "class MyClass:", "import numpy"]
144+
batch_tokens = vocab.tokenize(documents)
145+
146+
for doc, tokens in zip(documents, batch_tokens):
147+
print(f"{doc} -> {tokens}")
148+
```
149+
150+
### GPU Acceleration (if available)
151+
```python
152+
from crayon import CrayonVocab, check_backends
153+
154+
backends = check_backends()
155+
print(f"Available backends: {backends}")
156+
157+
if backends['cuda']:
158+
vocab = CrayonVocab(device="cuda")
159+
vocab.load_profile("science")
160+
161+
tokens = vocab.tokenize("E = mc²")
162+
print(tokens)
163+
```
164+
165+
---
166+
167+
## 📊 Benchmark Scripts
168+
169+
### Run Local Benchmarks
170+
```bash
171+
python local_benchmark.py
172+
```
173+
174+
### Run in Google Colab
175+
1. Open Google Colab
176+
2. Change runtime to GPU (T4/V100/A100)
177+
3. Copy contents of `Crayon_Colab_Notebook.py` or `colab_benchmark.py`
178+
4. Run the cell
179+
180+
---
181+
182+
## 🎉 Summary
183+
184+
XERV CRAYON v4.1.9 has been successfully:
185+
- ✅ Built with production-grade code
186+
- ✅ Tested on local hardware (64.6x faster than tiktoken)
187+
- ✅ Verified on Google Colab T4 GPU (10.2x faster than tiktoken)
188+
- ✅ Published to PyPI
189+
- ✅ Documented with comprehensive benchmarks
190+
- ✅ Ready for production use
191+
192+
**Install now:** `pip install xerv-crayon`
193+
194+
**View on PyPI:** https://pypi.org/project/xerv-crayon/4.1.9/

0 commit comments

Comments
 (0)