Spaces:
Running
Running
Commit
β’
bf397e6
1
Parent(s):
0f1bf97
changed optimizations layout and added them to plot
Browse files- app.py +13 -6
- src/assets/css_html_js.py +1 -2
app.py
CHANGED
@@ -29,14 +29,13 @@ COLUMNS_MAPPING = {
|
|
29 |
"model": "Model π€",
|
30 |
"backend.name": "Backend π",
|
31 |
"backend.torch_dtype": "Load Dtype π₯",
|
32 |
-
"
|
33 |
#
|
34 |
"forward.peak_memory(MB)": "Peak Memory (MB) β¬οΈ",
|
35 |
"generate.throughput(tokens/s)": "Throughput (tokens/s) β¬οΈ",
|
36 |
"average": "Average Open LLM Score β¬οΈ",
|
37 |
#
|
38 |
-
"
|
39 |
-
"backend.load_in_8bit": "LLM.int8 ποΈ",
|
40 |
}
|
41 |
COLUMNS_DATATYPES = [
|
42 |
"markdown",
|
@@ -49,7 +48,6 @@ COLUMNS_DATATYPES = [
|
|
49 |
"markdown",
|
50 |
#
|
51 |
"str",
|
52 |
-
"str",
|
53 |
]
|
54 |
SORTING_COLUMN = ["Throughput (tokens/s) β¬οΈ"]
|
55 |
|
@@ -66,6 +64,13 @@ def get_benchmark_df(benchmark="1xA100-80GB"):
|
|
66 |
scores_df = pd.read_csv(f"./llm-perf-dataset/reports/additional_data.csv")
|
67 |
bench_df = bench_df.merge(scores_df, on="model", how="left")
|
68 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
return bench_df
|
70 |
|
71 |
|
@@ -100,6 +105,7 @@ def get_benchmark_plot(bench_df):
|
|
100 |
"model",
|
101 |
"backend.name",
|
102 |
"backend.torch_dtype",
|
|
|
103 |
"forward.peak_memory(MB)",
|
104 |
"generate.throughput(tokens/s)",
|
105 |
],
|
@@ -129,8 +135,9 @@ def get_benchmark_plot(bench_df):
|
|
129 |
"Model: %{customdata[0]}",
|
130 |
"Backend: %{customdata[1]}",
|
131 |
"Datatype: %{customdata[2]}",
|
132 |
-
"
|
133 |
-
"
|
|
|
134 |
"Average Open LLM Score: %{y}",
|
135 |
"Per 1000 Tokens Latency (s): %{x}",
|
136 |
]
|
|
|
29 |
"model": "Model π€",
|
30 |
"backend.name": "Backend π",
|
31 |
"backend.torch_dtype": "Load Dtype π₯",
|
32 |
+
"optimizations": "Optimizations π οΈ",
|
33 |
#
|
34 |
"forward.peak_memory(MB)": "Peak Memory (MB) β¬οΈ",
|
35 |
"generate.throughput(tokens/s)": "Throughput (tokens/s) β¬οΈ",
|
36 |
"average": "Average Open LLM Score β¬οΈ",
|
37 |
#
|
38 |
+
"num_parameters": "#οΈβ£ Parameters π",
|
|
|
39 |
}
|
40 |
COLUMNS_DATATYPES = [
|
41 |
"markdown",
|
|
|
48 |
"markdown",
|
49 |
#
|
50 |
"str",
|
|
|
51 |
]
|
52 |
SORTING_COLUMN = ["Throughput (tokens/s) β¬οΈ"]
|
53 |
|
|
|
64 |
scores_df = pd.read_csv(f"./llm-perf-dataset/reports/additional_data.csv")
|
65 |
bench_df = bench_df.merge(scores_df, on="model", how="left")
|
66 |
|
67 |
+
bench_df["optimizations"] = bench_df[
|
68 |
+
["backend.bettertransformer", "backend.load_in_8bit"]
|
69 |
+
].apply(
|
70 |
+
lambda x: "BetterTransformer π€" if x[0] else "LLM.int8 ποΈ" if x[1] else "None",
|
71 |
+
axis=1,
|
72 |
+
)
|
73 |
+
|
74 |
return bench_df
|
75 |
|
76 |
|
|
|
105 |
"model",
|
106 |
"backend.name",
|
107 |
"backend.torch_dtype",
|
108 |
+
"optimizations",
|
109 |
"forward.peak_memory(MB)",
|
110 |
"generate.throughput(tokens/s)",
|
111 |
],
|
|
|
135 |
"Model: %{customdata[0]}",
|
136 |
"Backend: %{customdata[1]}",
|
137 |
"Datatype: %{customdata[2]}",
|
138 |
+
"Optimizations: %{customdata[3]}",
|
139 |
+
"Peak Memory (MB): %{customdata[4]}",
|
140 |
+
"Throughput (tokens/s): %{customdata[5]}",
|
141 |
"Average Open LLM Score: %{y}",
|
142 |
"Per 1000 Tokens Latency (s): %{x}",
|
143 |
]
|
src/assets/css_html_js.py
CHANGED
@@ -25,10 +25,9 @@ custom_css = """
|
|
25 |
border: none;
|
26 |
}
|
27 |
|
28 |
-
/* Limit the width of the first AutoEvalColumn so that names don't expand too much */
|
29 |
table td:first-child,
|
30 |
table th:first-child {
|
31 |
-
max-width:
|
32 |
overflow: auto;
|
33 |
white-space: nowrap;
|
34 |
}
|
|
|
25 |
border: none;
|
26 |
}
|
27 |
|
|
|
28 |
table td:first-child,
|
29 |
table th:first-child {
|
30 |
+
max-width: 300px;
|
31 |
overflow: auto;
|
32 |
white-space: nowrap;
|
33 |
}
|