Spaces:
Running
Running
Julien Simon
commited on
Commit
•
75e81c7
1
Parent(s):
6f23d6c
Update
Browse files- .gitignore +70 -0
- app.py +13 -15
- requirements.txt +1 -0
- results.py +101 -56
.gitignore
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Python
|
2 |
+
__pycache__/
|
3 |
+
*.py[cod]
|
4 |
+
*$py.class
|
5 |
+
*.so
|
6 |
+
.Python
|
7 |
+
build/
|
8 |
+
develop-eggs/
|
9 |
+
dist/
|
10 |
+
downloads/
|
11 |
+
eggs/
|
12 |
+
.eggs/
|
13 |
+
lib/
|
14 |
+
lib64/
|
15 |
+
parts/
|
16 |
+
sdist/
|
17 |
+
var/
|
18 |
+
wheels/
|
19 |
+
*.egg-info/
|
20 |
+
.installed.cfg
|
21 |
+
*.egg
|
22 |
+
|
23 |
+
# Jupyter Notebook
|
24 |
+
.ipynb_checkpoints
|
25 |
+
|
26 |
+
# IPython
|
27 |
+
profile_default/
|
28 |
+
ipython_config.py
|
29 |
+
|
30 |
+
# pyenv
|
31 |
+
.python-version
|
32 |
+
|
33 |
+
# Environments
|
34 |
+
.env
|
35 |
+
.venv
|
36 |
+
env/
|
37 |
+
venv/
|
38 |
+
ENV/
|
39 |
+
env.bak/
|
40 |
+
venv.bak/
|
41 |
+
|
42 |
+
# mypy
|
43 |
+
.mypy_cache/
|
44 |
+
.dmypy.json
|
45 |
+
dmypy.json
|
46 |
+
|
47 |
+
# Pytest
|
48 |
+
.pytest_cache/
|
49 |
+
|
50 |
+
# Coverage reports
|
51 |
+
htmlcov/
|
52 |
+
.coverage
|
53 |
+
.coverage.*
|
54 |
+
coverage.xml
|
55 |
+
*.cover
|
56 |
+
|
57 |
+
# VS Code
|
58 |
+
.vscode/
|
59 |
+
|
60 |
+
# PyCharm
|
61 |
+
.idea/
|
62 |
+
|
63 |
+
# Jupyter Book
|
64 |
+
_build/
|
65 |
+
|
66 |
+
# macOS
|
67 |
+
.DS_Store
|
68 |
+
|
69 |
+
# Windows
|
70 |
+
Thumbs.db
|
app.py
CHANGED
@@ -60,10 +60,6 @@ def display_results(model_name):
|
|
60 |
"""
|
61 |
Process and display results for a given model.
|
62 |
|
63 |
-
This function retrieves model data, processes it, and formats it for display.
|
64 |
-
It handles nested configurations, merges data from multiple models if necessary,
|
65 |
-
and sorts the results by instance type.
|
66 |
-
|
67 |
Args:
|
68 |
model_name (str): Name of the model to display results for.
|
69 |
|
@@ -86,10 +82,10 @@ def display_results(model_name):
|
|
86 |
merged_models.add(model.get('name', 'Unknown'))
|
87 |
for config in model.get('configurations', []):
|
88 |
try:
|
89 |
-
instance_type = config['instanceType']
|
90 |
cloud = config.get('cloud', 'N/A')
|
91 |
-
|
92 |
-
|
|
|
93 |
if 'configurations' in config:
|
94 |
for nested_config in config['configurations']:
|
95 |
nested_key = key + (nested_config.get('quantization', 'N/A'),)
|
@@ -100,24 +96,26 @@ def display_results(model_name):
|
|
100 |
"GPU RAM": config.get('gpuRAM', 'N/A'),
|
101 |
"Status": nested_config.get('status', 'N/A'),
|
102 |
"Quantization": nested_config.get('quantization', 'N/A'),
|
103 |
-
"
|
104 |
"Tokens per Second": nested_config.get('tokensPerSecond', 'N/A'),
|
105 |
-
"Notes": nested_config.get('notes', '')
|
106 |
}
|
107 |
else:
|
108 |
-
|
|
|
|
|
109 |
"Cloud": cloud,
|
110 |
"Instance Type": instance_type,
|
111 |
"GPU": config.get('gpu', 'N/A'),
|
112 |
"GPU RAM": config.get('gpuRAM', 'N/A'),
|
113 |
"Status": config.get('status', 'N/A'),
|
114 |
"Quantization": config.get('quantization', 'N/A'),
|
115 |
-
"
|
116 |
"Tokens per Second": config.get('tokensPerSecond', 'N/A'),
|
117 |
-
"Notes": config.get('notes', '')
|
118 |
}
|
119 |
-
except
|
120 |
-
|
121 |
continue
|
122 |
|
123 |
if not data:
|
@@ -165,7 +163,7 @@ def display_results(model_name):
|
|
165 |
|
166 |
with gr.Blocks() as demo:
|
167 |
gr.Markdown("# Model Benchmark Results")
|
168 |
-
gr.Markdown("This table shows the benchmark results for each model. [TGI](https://huggingface.co/docs/text-generation-inference/reference/launcher)
|
169 |
model_dropdown = gr.Dropdown(choices=get_model_names(), label="Select Model")
|
170 |
|
171 |
results_text = gr.Markdown()
|
|
|
60 |
"""
|
61 |
Process and display results for a given model.
|
62 |
|
|
|
|
|
|
|
|
|
63 |
Args:
|
64 |
model_name (str): Name of the model to display results for.
|
65 |
|
|
|
82 |
merged_models.add(model.get('name', 'Unknown'))
|
83 |
for config in model.get('configurations', []):
|
84 |
try:
|
|
|
85 |
cloud = config.get('cloud', 'N/A')
|
86 |
+
instance_type = config.get('instanceType', 'N/A')
|
87 |
+
key = (cloud, instance_type)
|
88 |
+
|
89 |
if 'configurations' in config:
|
90 |
for nested_config in config['configurations']:
|
91 |
nested_key = key + (nested_config.get('quantization', 'N/A'),)
|
|
|
96 |
"GPU RAM": config.get('gpuRAM', 'N/A'),
|
97 |
"Status": nested_config.get('status', 'N/A'),
|
98 |
"Quantization": nested_config.get('quantization', 'N/A'),
|
99 |
+
"Container": nested_config.get('container', nested_config.get('tgi', 'N/A')),
|
100 |
"Tokens per Second": nested_config.get('tokensPerSecond', 'N/A'),
|
101 |
+
"Notes": nested_config.get('notes', ''),
|
102 |
}
|
103 |
else:
|
104 |
+
# Generate a unique key for each configuration
|
105 |
+
unique_key = key + (config.get('quantization', 'N/A'), len(data))
|
106 |
+
data[unique_key] = {
|
107 |
"Cloud": cloud,
|
108 |
"Instance Type": instance_type,
|
109 |
"GPU": config.get('gpu', 'N/A'),
|
110 |
"GPU RAM": config.get('gpuRAM', 'N/A'),
|
111 |
"Status": config.get('status', 'N/A'),
|
112 |
"Quantization": config.get('quantization', 'N/A'),
|
113 |
+
"Container": config.get('container', config.get('tgi', 'N/A')),
|
114 |
"Tokens per Second": config.get('tokensPerSecond', 'N/A'),
|
115 |
+
"Notes": config.get('notes', ''),
|
116 |
}
|
117 |
+
except Exception as e:
|
118 |
+
print(f"Error processing configuration: {e}")
|
119 |
continue
|
120 |
|
121 |
if not data:
|
|
|
163 |
|
164 |
with gr.Blocks() as demo:
|
165 |
gr.Markdown("# Model Benchmark Results")
|
166 |
+
gr.Markdown("This table shows the benchmark results for each model. Container settings ([TGI](https://huggingface.co/docs/text-generation-inference/reference/launcher), [vLLM](https://docs.djl.ai/master/docs/serving/serving/docs/lmi/user_guides/vllm_user_guide.html), etc.) are default unless noted.")
|
167 |
model_dropdown = gr.Dropdown(choices=get_model_names(), label="Select Model")
|
168 |
|
169 |
results_text = gr.Markdown()
|
requirements.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
gradio
|
results.py
CHANGED
@@ -11,7 +11,7 @@ results = {
|
|
11 |
"gpu": "4xNVIDIA A10G",
|
12 |
"gpuRAM": "96 GB",
|
13 |
"quantization": "awq",
|
14 |
-
"
|
15 |
"status": "OK",
|
16 |
"tokensPerSecond": "33",
|
17 |
"notes": "",
|
@@ -23,7 +23,7 @@ results = {
|
|
23 |
"gpu": "4xNVIDIA A100",
|
24 |
"gpuRAM": "320 GB",
|
25 |
"quantization": "none",
|
26 |
-
"
|
27 |
"status": "OK",
|
28 |
"tokensPerSecond": "38",
|
29 |
"notes": "",
|
@@ -41,7 +41,7 @@ results = {
|
|
41 |
"gpu": "4xNVIDIA A10G",
|
42 |
"gpuRAM": "96 GB",
|
43 |
"quantization": "awq",
|
44 |
-
"
|
45 |
"status": "OK",
|
46 |
"tokensPerSecond": "33",
|
47 |
"notes": "",
|
@@ -53,7 +53,7 @@ results = {
|
|
53 |
"gpu": "4xNVIDIA A100",
|
54 |
"gpuRAM": "320 GB",
|
55 |
"quantization": "none",
|
56 |
-
"
|
57 |
"status": "OK",
|
58 |
"tokensPerSecond": "38",
|
59 |
"notes": "",
|
@@ -72,7 +72,7 @@ results = {
|
|
72 |
"gpu": "4xNVIDIA T4",
|
73 |
"gpuRAM": "64 GB",
|
74 |
"quantization": "bitsandbytes-nf4",
|
75 |
-
"
|
76 |
"status": "KO",
|
77 |
"tokensPerSecond": "-",
|
78 |
"notes": "Flash Attention requires Ampere GPUs or newer",
|
@@ -86,26 +86,26 @@ results = {
|
|
86 |
"configurations": [
|
87 |
{
|
88 |
"quantization": "bitsandbytes-nf4",
|
89 |
-
"
|
90 |
"status": "OK",
|
91 |
"tokensPerSecond": "12",
|
92 |
},
|
93 |
{
|
94 |
"quantization": "bitsandbytes-fp4",
|
95 |
-
"
|
96 |
"status": "OK",
|
97 |
"tokensPerSecond": "12",
|
98 |
},
|
99 |
{
|
100 |
"quantization": "bitsandbytes (int8)",
|
101 |
-
"
|
102 |
"status": "KO",
|
103 |
"tokensPerSecond": "-",
|
104 |
"notes": "CUDA OOM",
|
105 |
},
|
106 |
{
|
107 |
"quantization": "eetq (int8)",
|
108 |
-
"
|
109 |
"status": "KO",
|
110 |
"tokensPerSecond": "-",
|
111 |
"notes": "[FT Error] Heurisitc failed to find a valid config.",
|
@@ -121,26 +121,26 @@ results = {
|
|
121 |
"configurations": [
|
122 |
{
|
123 |
"quantization": "none",
|
124 |
-
"
|
125 |
"status": "KO",
|
126 |
"tokensPerSecond": "-",
|
127 |
"notes": "CUDA OOM (but g6.48xlarge works!)",
|
128 |
},
|
129 |
{
|
130 |
"quantization": "bitsandbytes-nf4",
|
131 |
-
"
|
132 |
"status": "OK",
|
133 |
"tokensPerSecond": "12.3",
|
134 |
},
|
135 |
{
|
136 |
"quantization": "bitsandbytes-fp4",
|
137 |
-
"
|
138 |
"status": "OK",
|
139 |
"tokensPerSecond": "12.5",
|
140 |
},
|
141 |
{
|
142 |
"quantization": "bitsandbytes (int8)",
|
143 |
-
"
|
144 |
"status": "KO",
|
145 |
"tokensPerSecond": "-",
|
146 |
"notes": "The model deploys, but inference times out.",
|
@@ -156,21 +156,21 @@ results = {
|
|
156 |
"configurations": [
|
157 |
{
|
158 |
"quantization": "bitsandbytes-nf4",
|
159 |
-
"
|
160 |
"status": "OK",
|
161 |
"tokensPerSecond": "1.5-2",
|
162 |
"notes": "Too slow, timeouts are likely",
|
163 |
},
|
164 |
{
|
165 |
"quantization": "bitsandbytes-fp4",
|
166 |
-
"
|
167 |
"status": "OK",
|
168 |
"tokensPerSecond": "2",
|
169 |
"notes": "Too slow, timeouts are likely",
|
170 |
},
|
171 |
{
|
172 |
"quantization": "bitsandbytes (int8)",
|
173 |
-
"
|
174 |
"status": "KO",
|
175 |
"tokensPerSecond": "-",
|
176 |
"notes": "CUDA OOM",
|
@@ -184,7 +184,7 @@ results = {
|
|
184 |
"gpu": "8xNVIDIA L4",
|
185 |
"gpuRAM": "192 GB",
|
186 |
"quantization": "none",
|
187 |
-
"
|
188 |
"status": "OK",
|
189 |
"tokensPerSecond": "12",
|
190 |
},
|
@@ -195,7 +195,7 @@ results = {
|
|
195 |
"gpu": "8xNVIDIA A100",
|
196 |
"gpuRAM": "320 GB",
|
197 |
"quantization": "none",
|
198 |
-
"
|
199 |
"status": "OK",
|
200 |
"tokensPerSecond": "40",
|
201 |
"notes": '"MAX_INPUT_LENGTH": "16384", "MAX_TOTAL_TOKENS": "32768",',
|
@@ -207,7 +207,7 @@ results = {
|
|
207 |
"gpu": "8xNVIDIA A100",
|
208 |
"gpuRAM": "320 GB",
|
209 |
"quantization": "none",
|
210 |
-
"
|
211 |
"status": "waiting for quota",
|
212 |
},
|
213 |
{
|
@@ -217,7 +217,7 @@ results = {
|
|
217 |
"gpu": "8xNVIDIA H100",
|
218 |
"gpuRAM": "640GB",
|
219 |
"quantization": "none",
|
220 |
-
"
|
221 |
"status": "OK",
|
222 |
"tokensPerSecond": "58",
|
223 |
"notes": '"MAX_INPUT_LENGTH": "16384", "MAX_TOTAL_TOKENS": "32768",',
|
@@ -227,7 +227,7 @@ results = {
|
|
227 |
"instanceType": "inf2.*",
|
228 |
"cloud": "AWS",
|
229 |
"gpu": "-",
|
230 |
-
"
|
231 |
"status": "not supported",
|
232 |
"tokensPerSecond": "-",
|
233 |
"notes": "Qwen2: TGI OK, Neuron SDK KO, optimum-neuron KO",
|
@@ -245,7 +245,7 @@ results = {
|
|
245 |
"gpu": "1xNVIDIA A10G",
|
246 |
"gpuRAM": "24 GB",
|
247 |
"quantization": "none",
|
248 |
-
"
|
249 |
"status": "OK",
|
250 |
"tokensPerSecond": "29",
|
251 |
"notes": "4K/8K fails",
|
@@ -257,7 +257,7 @@ results = {
|
|
257 |
"gpu": "4xNVIDIA A10G",
|
258 |
"gpuRAM": "96 GB",
|
259 |
"quantization": "none",
|
260 |
-
"
|
261 |
"status": "OK",
|
262 |
"tokensPerSecond": "85",
|
263 |
"notes": '"MAX_INPUT_TOKENS": "16384", "MAX_TOTAL_TOKENS": "32768",',
|
@@ -269,7 +269,7 @@ results = {
|
|
269 |
"gpu": "8xNVIDIA A10G",
|
270 |
"gpuRAM": "192 GB",
|
271 |
"quantization": "none",
|
272 |
-
"
|
273 |
"status": "OK",
|
274 |
"tokensPerSecond": "105",
|
275 |
"notes": '"MAX_INPUT_TOKENS": "20480", "MAX_TOTAL_TOKENS": "40960"\n\n32K/64K fails',
|
@@ -283,11 +283,11 @@ results = {
|
|
283 |
"configurations": [
|
284 |
{
|
285 |
"quantization": "none",
|
286 |
-
"
|
287 |
"status": "OK",
|
288 |
"tokensPerSecond": "15",
|
289 |
},
|
290 |
-
{"quantization": "fp8", "
|
291 |
],
|
292 |
},
|
293 |
{
|
@@ -297,7 +297,7 @@ results = {
|
|
297 |
"gpu": "4xNVIDIA L4",
|
298 |
"gpuRAM": "96 GB",
|
299 |
"quantization": "none",
|
300 |
-
"
|
301 |
"status": "OK",
|
302 |
"tokensPerSecond": "51",
|
303 |
"notes": "same as g5?",
|
@@ -309,7 +309,7 @@ results = {
|
|
309 |
"gpu": "8xNVIDIA L4",
|
310 |
"gpuRAM": "192 GB",
|
311 |
"quantization": "none",
|
312 |
-
"
|
313 |
"status": "OK",
|
314 |
"tokensPerSecond": "81",
|
315 |
"notes": "same as g5?",
|
@@ -321,7 +321,7 @@ results = {
|
|
321 |
"gpu": "1xNVIDIA L40S",
|
322 |
"gpuRAM": "48 GB",
|
323 |
"quantization": "none",
|
324 |
-
"
|
325 |
"status": "OK",
|
326 |
"tokensPerSecond": "42.1",
|
327 |
},
|
@@ -332,9 +332,20 @@ results = {
|
|
332 |
"gpu": "1xNVIDIA L40S",
|
333 |
"gpuRAM": "48 GB",
|
334 |
"quantization": "none",
|
335 |
-
"
|
336 |
"status": "OK",
|
337 |
-
"tokensPerSecond": "45"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
338 |
},
|
339 |
{
|
340 |
"region": "AWS",
|
@@ -343,7 +354,7 @@ results = {
|
|
343 |
"gpu": "4xNVIDIA A100",
|
344 |
"gpuRAM": "320 GB",
|
345 |
"quantization": "none",
|
346 |
-
"
|
347 |
"status": "OK",
|
348 |
"tokensPerSecond": "145",
|
349 |
"notes": '"MAX_INPUT_TOKENS": "40960", "MAX_TOTAL_TOKENS": "81920"\n\n64K/128K fails (even with 4-bit)',
|
@@ -371,7 +382,7 @@ results = {
|
|
371 |
"gpu": "1xNVIDIA A10G",
|
372 |
"gpuRAM": "24 GB",
|
373 |
"quantization": "none",
|
374 |
-
"
|
375 |
"status": "OK",
|
376 |
"tokensPerSecond": "30",
|
377 |
},
|
@@ -382,7 +393,7 @@ results = {
|
|
382 |
"gpu": "4xNVIDIA A10G",
|
383 |
"gpuRAM": "96 GB",
|
384 |
"quantization": "none",
|
385 |
-
"
|
386 |
"status": "OK",
|
387 |
"tokensPerSecond": "83",
|
388 |
},
|
@@ -393,7 +404,7 @@ results = {
|
|
393 |
"gpu": "8xNVIDIA A10G",
|
394 |
"gpuRAM": "192 GB",
|
395 |
"quantization": "none",
|
396 |
-
"
|
397 |
"status": "KO",
|
398 |
"tokensPerSecond": "-",
|
399 |
"notes": "ValueError: `num_heads` must be divisible by `num_shards` (got `num_heads`: 28 and `num_shards`: 8\n\nSM_NUM_GPUS=7 doesn't work either because tensor size ares not a multiple of 7 (e.g., 512)",
|
@@ -405,7 +416,7 @@ results = {
|
|
405 |
"gpu": "1xNVIDIA L4",
|
406 |
"gpuRAM": "24 GB",
|
407 |
"quantization": "none",
|
408 |
-
"
|
409 |
"status": "OK",
|
410 |
"tokensPerSecond": "16.3",
|
411 |
},
|
@@ -416,7 +427,7 @@ results = {
|
|
416 |
"gpu": "4xNVIDIA L4",
|
417 |
"gpuRAM": "96 GB",
|
418 |
"quantization": "none",
|
419 |
-
"
|
420 |
"status": "OK",
|
421 |
"tokensPerSecond": "54.2",
|
422 |
},
|
@@ -425,14 +436,48 @@ results = {
|
|
425 |
"instanceType": "inf2.*",
|
426 |
"cloud": "AWS",
|
427 |
"gpu": "-",
|
428 |
-
"
|
429 |
"status": "not supported",
|
430 |
"tokensPerSecond": "-",
|
431 |
"notes": "Qwen2: TGI OK, Neuron SDK KO, optimum-neuron KO",
|
432 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
433 |
],
|
434 |
},
|
435 |
-
{"name": "Arcee-Spark",
|
|
|
436 |
{
|
437 |
"name": "Arcee-Lite",
|
438 |
"modelType": "Qwen2 1.5B distilled from phi-3-medium 14B",
|
@@ -444,7 +489,7 @@ results = {
|
|
444 |
"gpu": "-",
|
445 |
"gpuRAM": "-",
|
446 |
"quantization": "bitsandbytes-nf4",
|
447 |
-
"
|
448 |
"status": "KO",
|
449 |
"tokensPerSecond": "-",
|
450 |
"notes": "OOM, might work with a prequantized model",
|
@@ -456,7 +501,7 @@ results = {
|
|
456 |
"gpu": "-",
|
457 |
"gpuRAM": "-",
|
458 |
"quantization": "bitsandbytes-nf4",
|
459 |
-
"
|
460 |
"status": "KO",
|
461 |
"tokensPerSecond": "-",
|
462 |
"notes": "OOM, might work with a prequantized model",
|
@@ -470,19 +515,19 @@ results = {
|
|
470 |
"configurations": [
|
471 |
{
|
472 |
"quantization": "none",
|
473 |
-
"
|
474 |
"status": "OK",
|
475 |
"tokensPerSecond": "10.7",
|
476 |
},
|
477 |
{
|
478 |
"quantization": "bitsandbytes (int8)",
|
479 |
-
"
|
480 |
"status": "OK",
|
481 |
"tokensPerSecond": "10.5",
|
482 |
},
|
483 |
{
|
484 |
"quantization": "bitsandbytes-nf4",
|
485 |
-
"
|
486 |
"status": "OK",
|
487 |
"tokensPerSecond": "10.6",
|
488 |
},
|
@@ -495,7 +540,7 @@ results = {
|
|
495 |
"gpu": "-",
|
496 |
"gpuRAM": "-",
|
497 |
"quantization": "none",
|
498 |
-
"
|
499 |
"status": "waiting for quota",
|
500 |
"tokensPerSecond": "-",
|
501 |
},
|
@@ -508,13 +553,13 @@ results = {
|
|
508 |
"configurations": [
|
509 |
{
|
510 |
"quantization": "none",
|
511 |
-
"
|
512 |
"status": "OK",
|
513 |
"tokensPerSecond": "110",
|
514 |
},
|
515 |
{
|
516 |
"quantization": "none",
|
517 |
-
"
|
518 |
"status": "OK",
|
519 |
"tokensPerSecond": "105",
|
520 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",',
|
@@ -528,7 +573,7 @@ results = {
|
|
528 |
"gpu": "1xNVIDIA L40S",
|
529 |
"gpuRAM": "48 GB",
|
530 |
"quantization": "none",
|
531 |
-
"
|
532 |
"status": "OK",
|
533 |
"tokensPerSecond": "160",
|
534 |
},
|
@@ -544,7 +589,7 @@ results = {
|
|
544 |
"gpu": "1xNVIDIA A10G",
|
545 |
"gpuRAM": "24 GB",
|
546 |
"quantization": "none",
|
547 |
-
"
|
548 |
"status": "OK",
|
549 |
"tokensPerSecond": 29,
|
550 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
@@ -555,7 +600,7 @@ results = {
|
|
555 |
"gpu": "4xNVIDIA A10G",
|
556 |
"gpuRAM": "96 GB",
|
557 |
"quantization": "none",
|
558 |
-
"
|
559 |
"status": "OK",
|
560 |
"tokensPerSecond": 65,
|
561 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",\nNot supported by AutoAWQ and AutoGPTQ',
|
@@ -566,7 +611,7 @@ results = {
|
|
566 |
"gpu": "8xNVIDIA A10G",
|
567 |
"gpuRAM": "192 GB",
|
568 |
"quantization": "none",
|
569 |
-
"
|
570 |
"status": "OK",
|
571 |
"tokensPerSecond": 80,
|
572 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
@@ -577,7 +622,7 @@ results = {
|
|
577 |
"gpu": "1xNVIDIA L4",
|
578 |
"gpuRAM": "24 GB",
|
579 |
"quantization": "none",
|
580 |
-
"
|
581 |
"status": "OK",
|
582 |
"tokensPerSecond": 16,
|
583 |
"notes": '"OPTION_MAX_MODEL_LEN": "4096"',
|
@@ -588,7 +633,7 @@ results = {
|
|
588 |
"gpu": "4xNVIDIA L4",
|
589 |
"gpuRAM": "96 GB",
|
590 |
"quantization": "none",
|
591 |
-
"
|
592 |
"status": "OK",
|
593 |
"tokensPerSecond": 50,
|
594 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
@@ -599,7 +644,7 @@ results = {
|
|
599 |
"gpu": "8xNVIDIA L4",
|
600 |
"gpuRAM": "192 GB",
|
601 |
"quantization": "none",
|
602 |
-
"
|
603 |
"status": "OK",
|
604 |
"tokensPerSecond": 69,
|
605 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
@@ -610,7 +655,7 @@ results = {
|
|
610 |
"gpu": "1xNVIDIA L40S",
|
611 |
"gpuRAM": "48 GB",
|
612 |
"quantization": "none",
|
613 |
-
"
|
614 |
"status": "OK",
|
615 |
"tokensPerSecond": 46,
|
616 |
},
|
@@ -620,7 +665,7 @@ results = {
|
|
620 |
"gpu": "4xNVIDIA A100",
|
621 |
"gpuRAM": "320 GB",
|
622 |
"quantization": "none",
|
623 |
-
"
|
624 |
"status": "OK",
|
625 |
"tokensPerSecond": 82,
|
626 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
|
|
11 |
"gpu": "4xNVIDIA A10G",
|
12 |
"gpuRAM": "96 GB",
|
13 |
"quantization": "awq",
|
14 |
+
"container": "TGI 2.2.0",
|
15 |
"status": "OK",
|
16 |
"tokensPerSecond": "33",
|
17 |
"notes": "",
|
|
|
23 |
"gpu": "4xNVIDIA A100",
|
24 |
"gpuRAM": "320 GB",
|
25 |
"quantization": "none",
|
26 |
+
"container": "TGI 2.2.0",
|
27 |
"status": "OK",
|
28 |
"tokensPerSecond": "38",
|
29 |
"notes": "",
|
|
|
41 |
"gpu": "4xNVIDIA A10G",
|
42 |
"gpuRAM": "96 GB",
|
43 |
"quantization": "awq",
|
44 |
+
"container": "TGI 2.2.0",
|
45 |
"status": "OK",
|
46 |
"tokensPerSecond": "33",
|
47 |
"notes": "",
|
|
|
53 |
"gpu": "4xNVIDIA A100",
|
54 |
"gpuRAM": "320 GB",
|
55 |
"quantization": "none",
|
56 |
+
"container": "TGI 2.2.0",
|
57 |
"status": "OK",
|
58 |
"tokensPerSecond": "38",
|
59 |
"notes": "",
|
|
|
72 |
"gpu": "4xNVIDIA T4",
|
73 |
"gpuRAM": "64 GB",
|
74 |
"quantization": "bitsandbytes-nf4",
|
75 |
+
"container": "TGI 2.2.0",
|
76 |
"status": "KO",
|
77 |
"tokensPerSecond": "-",
|
78 |
"notes": "Flash Attention requires Ampere GPUs or newer",
|
|
|
86 |
"configurations": [
|
87 |
{
|
88 |
"quantization": "bitsandbytes-nf4",
|
89 |
+
"container": "TGI 2.2.0",
|
90 |
"status": "OK",
|
91 |
"tokensPerSecond": "12",
|
92 |
},
|
93 |
{
|
94 |
"quantization": "bitsandbytes-fp4",
|
95 |
+
"container": "TGI 2.2.0",
|
96 |
"status": "OK",
|
97 |
"tokensPerSecond": "12",
|
98 |
},
|
99 |
{
|
100 |
"quantization": "bitsandbytes (int8)",
|
101 |
+
"container": "TGI 2.2.0",
|
102 |
"status": "KO",
|
103 |
"tokensPerSecond": "-",
|
104 |
"notes": "CUDA OOM",
|
105 |
},
|
106 |
{
|
107 |
"quantization": "eetq (int8)",
|
108 |
+
"container": "TGI 2.2.0",
|
109 |
"status": "KO",
|
110 |
"tokensPerSecond": "-",
|
111 |
"notes": "[FT Error] Heurisitc failed to find a valid config.",
|
|
|
121 |
"configurations": [
|
122 |
{
|
123 |
"quantization": "none",
|
124 |
+
"container": "TGI 2.2.0",
|
125 |
"status": "KO",
|
126 |
"tokensPerSecond": "-",
|
127 |
"notes": "CUDA OOM (but g6.48xlarge works!)",
|
128 |
},
|
129 |
{
|
130 |
"quantization": "bitsandbytes-nf4",
|
131 |
+
"container": "TGI 2.2.0",
|
132 |
"status": "OK",
|
133 |
"tokensPerSecond": "12.3",
|
134 |
},
|
135 |
{
|
136 |
"quantization": "bitsandbytes-fp4",
|
137 |
+
"container": "TGI 2.2.0",
|
138 |
"status": "OK",
|
139 |
"tokensPerSecond": "12.5",
|
140 |
},
|
141 |
{
|
142 |
"quantization": "bitsandbytes (int8)",
|
143 |
+
"container": "TGI 2.2.0",
|
144 |
"status": "KO",
|
145 |
"tokensPerSecond": "-",
|
146 |
"notes": "The model deploys, but inference times out.",
|
|
|
156 |
"configurations": [
|
157 |
{
|
158 |
"quantization": "bitsandbytes-nf4",
|
159 |
+
"container": "TGI 2.2.0",
|
160 |
"status": "OK",
|
161 |
"tokensPerSecond": "1.5-2",
|
162 |
"notes": "Too slow, timeouts are likely",
|
163 |
},
|
164 |
{
|
165 |
"quantization": "bitsandbytes-fp4",
|
166 |
+
"container": "TGI 2.2.0",
|
167 |
"status": "OK",
|
168 |
"tokensPerSecond": "2",
|
169 |
"notes": "Too slow, timeouts are likely",
|
170 |
},
|
171 |
{
|
172 |
"quantization": "bitsandbytes (int8)",
|
173 |
+
"container": "TGI 2.2.0",
|
174 |
"status": "KO",
|
175 |
"tokensPerSecond": "-",
|
176 |
"notes": "CUDA OOM",
|
|
|
184 |
"gpu": "8xNVIDIA L4",
|
185 |
"gpuRAM": "192 GB",
|
186 |
"quantization": "none",
|
187 |
+
"container": "TGI 2.2.0",
|
188 |
"status": "OK",
|
189 |
"tokensPerSecond": "12",
|
190 |
},
|
|
|
195 |
"gpu": "8xNVIDIA A100",
|
196 |
"gpuRAM": "320 GB",
|
197 |
"quantization": "none",
|
198 |
+
"container": "TGI 2.2.0",
|
199 |
"status": "OK",
|
200 |
"tokensPerSecond": "40",
|
201 |
"notes": '"MAX_INPUT_LENGTH": "16384", "MAX_TOTAL_TOKENS": "32768",',
|
|
|
207 |
"gpu": "8xNVIDIA A100",
|
208 |
"gpuRAM": "320 GB",
|
209 |
"quantization": "none",
|
210 |
+
"container": "TGI 2.2.0",
|
211 |
"status": "waiting for quota",
|
212 |
},
|
213 |
{
|
|
|
217 |
"gpu": "8xNVIDIA H100",
|
218 |
"gpuRAM": "640GB",
|
219 |
"quantization": "none",
|
220 |
+
"container": "TGI 2.2.0",
|
221 |
"status": "OK",
|
222 |
"tokensPerSecond": "58",
|
223 |
"notes": '"MAX_INPUT_LENGTH": "16384", "MAX_TOTAL_TOKENS": "32768",',
|
|
|
227 |
"instanceType": "inf2.*",
|
228 |
"cloud": "AWS",
|
229 |
"gpu": "-",
|
230 |
+
"container": "TGI 2.2.0",
|
231 |
"status": "not supported",
|
232 |
"tokensPerSecond": "-",
|
233 |
"notes": "Qwen2: TGI OK, Neuron SDK KO, optimum-neuron KO",
|
|
|
245 |
"gpu": "1xNVIDIA A10G",
|
246 |
"gpuRAM": "24 GB",
|
247 |
"quantization": "none",
|
248 |
+
"container": "TGI 2.2.0",
|
249 |
"status": "OK",
|
250 |
"tokensPerSecond": "29",
|
251 |
"notes": "4K/8K fails",
|
|
|
257 |
"gpu": "4xNVIDIA A10G",
|
258 |
"gpuRAM": "96 GB",
|
259 |
"quantization": "none",
|
260 |
+
"container": "TGI 2.2.0",
|
261 |
"status": "OK",
|
262 |
"tokensPerSecond": "85",
|
263 |
"notes": '"MAX_INPUT_TOKENS": "16384", "MAX_TOTAL_TOKENS": "32768",',
|
|
|
269 |
"gpu": "8xNVIDIA A10G",
|
270 |
"gpuRAM": "192 GB",
|
271 |
"quantization": "none",
|
272 |
+
"container": "TGI 2.2.0",
|
273 |
"status": "OK",
|
274 |
"tokensPerSecond": "105",
|
275 |
"notes": '"MAX_INPUT_TOKENS": "20480", "MAX_TOTAL_TOKENS": "40960"\n\n32K/64K fails',
|
|
|
283 |
"configurations": [
|
284 |
{
|
285 |
"quantization": "none",
|
286 |
+
"container": "TGI 2.2.0",
|
287 |
"status": "OK",
|
288 |
"tokensPerSecond": "15",
|
289 |
},
|
290 |
+
{"quantization": "fp8", "container": "TGI 2.2.0"},
|
291 |
],
|
292 |
},
|
293 |
{
|
|
|
297 |
"gpu": "4xNVIDIA L4",
|
298 |
"gpuRAM": "96 GB",
|
299 |
"quantization": "none",
|
300 |
+
"container": "TGI 2.2.0",
|
301 |
"status": "OK",
|
302 |
"tokensPerSecond": "51",
|
303 |
"notes": "same as g5?",
|
|
|
309 |
"gpu": "8xNVIDIA L4",
|
310 |
"gpuRAM": "192 GB",
|
311 |
"quantization": "none",
|
312 |
+
"container": "TGI 2.2.0",
|
313 |
"status": "OK",
|
314 |
"tokensPerSecond": "81",
|
315 |
"notes": "same as g5?",
|
|
|
321 |
"gpu": "1xNVIDIA L40S",
|
322 |
"gpuRAM": "48 GB",
|
323 |
"quantization": "none",
|
324 |
+
"container": "TGI 2.2.0",
|
325 |
"status": "OK",
|
326 |
"tokensPerSecond": "42.1",
|
327 |
},
|
|
|
332 |
"gpu": "1xNVIDIA L40S",
|
333 |
"gpuRAM": "48 GB",
|
334 |
"quantization": "none",
|
335 |
+
"container": "SGLang 0.2.13",
|
336 |
"status": "OK",
|
337 |
+
"tokensPerSecond": "45"
|
338 |
+
},
|
339 |
+
{
|
340 |
+
"region": "AWS",
|
341 |
+
"instanceType": "g6e.2xlarge",
|
342 |
+
"cloud": "AWS",
|
343 |
+
"gpu": "1xNVIDIA L40S",
|
344 |
+
"gpuRAM": "48 GB",
|
345 |
+
"quantization": "none",
|
346 |
+
"container": "vLLM 0.5.5",
|
347 |
+
"status": "OK",
|
348 |
+
"tokensPerSecond": "43.4"
|
349 |
},
|
350 |
{
|
351 |
"region": "AWS",
|
|
|
354 |
"gpu": "4xNVIDIA A100",
|
355 |
"gpuRAM": "320 GB",
|
356 |
"quantization": "none",
|
357 |
+
"container": "TGI 2.2.0",
|
358 |
"status": "OK",
|
359 |
"tokensPerSecond": "145",
|
360 |
"notes": '"MAX_INPUT_TOKENS": "40960", "MAX_TOTAL_TOKENS": "81920"\n\n64K/128K fails (even with 4-bit)',
|
|
|
382 |
"gpu": "1xNVIDIA A10G",
|
383 |
"gpuRAM": "24 GB",
|
384 |
"quantization": "none",
|
385 |
+
"container": "TGI 2.2.0",
|
386 |
"status": "OK",
|
387 |
"tokensPerSecond": "30",
|
388 |
},
|
|
|
393 |
"gpu": "4xNVIDIA A10G",
|
394 |
"gpuRAM": "96 GB",
|
395 |
"quantization": "none",
|
396 |
+
"container": "TGI 2.2.0",
|
397 |
"status": "OK",
|
398 |
"tokensPerSecond": "83",
|
399 |
},
|
|
|
404 |
"gpu": "8xNVIDIA A10G",
|
405 |
"gpuRAM": "192 GB",
|
406 |
"quantization": "none",
|
407 |
+
"container": "TGI 2.2.0",
|
408 |
"status": "KO",
|
409 |
"tokensPerSecond": "-",
|
410 |
"notes": "ValueError: `num_heads` must be divisible by `num_shards` (got `num_heads`: 28 and `num_shards`: 8\n\nSM_NUM_GPUS=7 doesn't work either because tensor size ares not a multiple of 7 (e.g., 512)",
|
|
|
416 |
"gpu": "1xNVIDIA L4",
|
417 |
"gpuRAM": "24 GB",
|
418 |
"quantization": "none",
|
419 |
+
"container": "TGI 2.2.0",
|
420 |
"status": "OK",
|
421 |
"tokensPerSecond": "16.3",
|
422 |
},
|
|
|
427 |
"gpu": "4xNVIDIA L4",
|
428 |
"gpuRAM": "96 GB",
|
429 |
"quantization": "none",
|
430 |
+
"container": "TGI 2.2.0",
|
431 |
"status": "OK",
|
432 |
"tokensPerSecond": "54.2",
|
433 |
},
|
|
|
436 |
"instanceType": "inf2.*",
|
437 |
"cloud": "AWS",
|
438 |
"gpu": "-",
|
439 |
+
"container": "TGI 2.2.0",
|
440 |
"status": "not supported",
|
441 |
"tokensPerSecond": "-",
|
442 |
"notes": "Qwen2: TGI OK, Neuron SDK KO, optimum-neuron KO",
|
443 |
},
|
444 |
+
{
|
445 |
+
"region": "us-west-2",
|
446 |
+
"instanceType": "g6e.2xlarge",
|
447 |
+
"cloud": "AWS",
|
448 |
+
"gpu": "1xNVIDIA L40S",
|
449 |
+
"gpuRAM": "48 GB",
|
450 |
+
"quantization": "none",
|
451 |
+
"container": "TGI 2.2.0",
|
452 |
+
"status": "OK",
|
453 |
+
"tokensPerSecond": "45",
|
454 |
+
},
|
455 |
+
{
|
456 |
+
"region": "us-west-2",
|
457 |
+
"instanceType": "g6e.2xlarge",
|
458 |
+
"cloud": "AWS",
|
459 |
+
"gpu": "1xNVIDIA L40S",
|
460 |
+
"gpuRAM": "48 GB",
|
461 |
+
"quantization": "none",
|
462 |
+
"container": "SGLang 0.2.13",
|
463 |
+
"status": "OK",
|
464 |
+
"tokensPerSecond": "48",
|
465 |
+
},
|
466 |
+
{
|
467 |
+
"region": "us-west-2",
|
468 |
+
"instanceType": "g6e.2xlarge",
|
469 |
+
"cloud": "AWS",
|
470 |
+
"gpu": "1xNVIDIA L40S",
|
471 |
+
"gpuRAM": "48 GB",
|
472 |
+
"quantization": "none",
|
473 |
+
"container": "vLLM 0.5.5",
|
474 |
+
"status": "OK",
|
475 |
+
"tokensPerSecond": "45.7",
|
476 |
+
},
|
477 |
],
|
478 |
},
|
479 |
+
{"name": "Arcee-Spark",
|
480 |
+
"modelType": "Qwen2 7B"},
|
481 |
{
|
482 |
"name": "Arcee-Lite",
|
483 |
"modelType": "Qwen2 1.5B distilled from phi-3-medium 14B",
|
|
|
489 |
"gpu": "-",
|
490 |
"gpuRAM": "-",
|
491 |
"quantization": "bitsandbytes-nf4",
|
492 |
+
"container": "TGI 2.2.0",
|
493 |
"status": "KO",
|
494 |
"tokensPerSecond": "-",
|
495 |
"notes": "OOM, might work with a prequantized model",
|
|
|
501 |
"gpu": "-",
|
502 |
"gpuRAM": "-",
|
503 |
"quantization": "bitsandbytes-nf4",
|
504 |
+
"container": "TGI 2.2.0",
|
505 |
"status": "KO",
|
506 |
"tokensPerSecond": "-",
|
507 |
"notes": "OOM, might work with a prequantized model",
|
|
|
515 |
"configurations": [
|
516 |
{
|
517 |
"quantization": "none",
|
518 |
+
"container": "TGI 2.2.0",
|
519 |
"status": "OK",
|
520 |
"tokensPerSecond": "10.7",
|
521 |
},
|
522 |
{
|
523 |
"quantization": "bitsandbytes (int8)",
|
524 |
+
"container": "TGI 2.2.0",
|
525 |
"status": "OK",
|
526 |
"tokensPerSecond": "10.5",
|
527 |
},
|
528 |
{
|
529 |
"quantization": "bitsandbytes-nf4",
|
530 |
+
"container": "TGI 2.2.0",
|
531 |
"status": "OK",
|
532 |
"tokensPerSecond": "10.6",
|
533 |
},
|
|
|
540 |
"gpu": "-",
|
541 |
"gpuRAM": "-",
|
542 |
"quantization": "none",
|
543 |
+
"container": "TGI 2.2.0",
|
544 |
"status": "waiting for quota",
|
545 |
"tokensPerSecond": "-",
|
546 |
},
|
|
|
553 |
"configurations": [
|
554 |
{
|
555 |
"quantization": "none",
|
556 |
+
"container": "TGI 2.2.0",
|
557 |
"status": "OK",
|
558 |
"tokensPerSecond": "110",
|
559 |
},
|
560 |
{
|
561 |
"quantization": "none",
|
562 |
+
"container": "DJL 0.28 vLLM",
|
563 |
"status": "OK",
|
564 |
"tokensPerSecond": "105",
|
565 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",',
|
|
|
573 |
"gpu": "1xNVIDIA L40S",
|
574 |
"gpuRAM": "48 GB",
|
575 |
"quantization": "none",
|
576 |
+
"container": "TGI 2.2.0",
|
577 |
"status": "OK",
|
578 |
"tokensPerSecond": "160",
|
579 |
},
|
|
|
589 |
"gpu": "1xNVIDIA A10G",
|
590 |
"gpuRAM": "24 GB",
|
591 |
"quantization": "none",
|
592 |
+
"container": "DJL 0.28 vLLM",
|
593 |
"status": "OK",
|
594 |
"tokensPerSecond": 29,
|
595 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
|
|
600 |
"gpu": "4xNVIDIA A10G",
|
601 |
"gpuRAM": "96 GB",
|
602 |
"quantization": "none",
|
603 |
+
"container": "DJL 0.28 vLLM",
|
604 |
"status": "OK",
|
605 |
"tokensPerSecond": 65,
|
606 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",\nNot supported by AutoAWQ and AutoGPTQ',
|
|
|
611 |
"gpu": "8xNVIDIA A10G",
|
612 |
"gpuRAM": "192 GB",
|
613 |
"quantization": "none",
|
614 |
+
"container": "DJL 0.28 vLLM",
|
615 |
"status": "OK",
|
616 |
"tokensPerSecond": 80,
|
617 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
|
|
622 |
"gpu": "1xNVIDIA L4",
|
623 |
"gpuRAM": "24 GB",
|
624 |
"quantization": "none",
|
625 |
+
"container": "DJL 0.28 vLLM",
|
626 |
"status": "OK",
|
627 |
"tokensPerSecond": 16,
|
628 |
"notes": '"OPTION_MAX_MODEL_LEN": "4096"',
|
|
|
633 |
"gpu": "4xNVIDIA L4",
|
634 |
"gpuRAM": "96 GB",
|
635 |
"quantization": "none",
|
636 |
+
"container": "DJL 0.28 vLLM",
|
637 |
"status": "OK",
|
638 |
"tokensPerSecond": 50,
|
639 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
|
|
644 |
"gpu": "8xNVIDIA L4",
|
645 |
"gpuRAM": "192 GB",
|
646 |
"quantization": "none",
|
647 |
+
"container": "DJL 0.28 vLLM",
|
648 |
"status": "OK",
|
649 |
"tokensPerSecond": 69,
|
650 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|
|
|
655 |
"gpu": "1xNVIDIA L40S",
|
656 |
"gpuRAM": "48 GB",
|
657 |
"quantization": "none",
|
658 |
+
"container": "SGLang 0.2.13",
|
659 |
"status": "OK",
|
660 |
"tokensPerSecond": 46,
|
661 |
},
|
|
|
665 |
"gpu": "4xNVIDIA A100",
|
666 |
"gpuRAM": "320 GB",
|
667 |
"quantization": "none",
|
668 |
+
"container": "DJL 0.28 vLLM",
|
669 |
"status": "OK",
|
670 |
"tokensPerSecond": 82,
|
671 |
"notes": '"OPTION_MAX_MODEL_LEN": "32768",\n"TENSOR_PARALLEL_DEGREE": "max",',
|