make table more clear- again
Browse files
README.md
CHANGED
@@ -8,28 +8,28 @@ Note:
|
|
8 |
|
9 |
- The "bfl codebase" refers to the original [BFL codebase](https://github.com/black-forest-labs/flux), not this repo.
|
10 |
- The "fp8 wo quant" refers to the original BFL codebase using fp8 weight only quantization, not using fp8 matmul which is default in this repo.
|
11 |
-
- The "compile blocks & extras" refers to the option within this repo setting the config values `"compile_blocks" true` & `"compiled_extras": true`.
|
12 |
-
- All generations which including a
|
13 |
|
14 |
| Resolution | Device | Test | Average it/s |
|
15 |
| ---------- | ---------- | -------------------------- | ------------ |
|
16 |
| 1024x1024 | RTX4090 | bfl codebase fp8 wo quant | 1.7 |
|
17 |
-
| 1024x1024 | RTX4090 |
|
18 |
| 1024x1024 | RTX4090 | β
compile blocks & extras | 3.51 |
|
19 |
| 1024x1024 | RTX6000ADA | bfl codebase | 1.74 |
|
20 |
-
| 1024x1024 | RTX6000ADA |
|
21 |
| 1024x1024 | RTX6000ADA | β
compile blocks & extras | 2.8 |
|
22 |
| 768x768 | RTX4090 | bfl codebase fp8 wo quant | 2.32 |
|
23 |
-
| 768x768 | RTX4090 |
|
24 |
| 768x768 | RTX4090 | β
compile blocks & extras | 6.2 |
|
25 |
| 768x768 | RTX6000ADA | bfl codebase | 3.01 |
|
26 |
-
| 768x768 | RTX6000ADA |
|
27 |
| 768x768 | RTX6000ADA | β
compile blocks & extras | 4.46 |
|
28 |
| 1024x720 | RTX4090 | bfl codebase fp8 wo quant | 3.01 |
|
29 |
-
| 1024x720 | RTX4090 |
|
30 |
| 1024x720 | RTX4090 | β
compile blocks & extras | 4.96 |
|
31 |
| 1024x720 | RTX6000ADA | bfl codebase | 2.37 |
|
32 |
-
| 1024x720 | RTX6000ADA |
|
33 |
| 1024x720 | RTX6000ADA | β
compile blocks & extras | 3.78 |
|
34 |
|
35 |
## Table of Contents
|
|
|
8 |
|
9 |
- The "bfl codebase" refers to the original [BFL codebase](https://github.com/black-forest-labs/flux), not this repo.
|
10 |
- The "fp8 wo quant" refers to the original BFL codebase using fp8 weight only quantization, not using fp8 matmul which is default in this repo.
|
11 |
+
- The "compile blocks & extras" refers to the option within this repo setting the config values `"compile_blocks" true` & `"compiled_extras": true`. β means both were set to false, β
means both were set to true.
|
12 |
+
- All generations which including a β or β
are using this repo.
|
13 |
|
14 |
| Resolution | Device | Test | Average it/s |
|
15 |
| ---------- | ---------- | -------------------------- | ------------ |
|
16 |
| 1024x1024 | RTX4090 | bfl codebase fp8 wo quant | 1.7 |
|
17 |
+
| 1024x1024 | RTX4090 | β compile blocks & extras | 2.55 |
|
18 |
| 1024x1024 | RTX4090 | β
compile blocks & extras | 3.51 |
|
19 |
| 1024x1024 | RTX6000ADA | bfl codebase | 1.74 |
|
20 |
+
| 1024x1024 | RTX6000ADA | β compile blocks & extras | 2.08 |
|
21 |
| 1024x1024 | RTX6000ADA | β
compile blocks & extras | 2.8 |
|
22 |
| 768x768 | RTX4090 | bfl codebase fp8 wo quant | 2.32 |
|
23 |
+
| 768x768 | RTX4090 | β compile blocks & extras | 4.47 |
|
24 |
| 768x768 | RTX4090 | β
compile blocks & extras | 6.2 |
|
25 |
| 768x768 | RTX6000ADA | bfl codebase | 3.01 |
|
26 |
+
| 768x768 | RTX6000ADA | β compile blocks & extras | 3.43 |
|
27 |
| 768x768 | RTX6000ADA | β
compile blocks & extras | 4.46 |
|
28 |
| 1024x720 | RTX4090 | bfl codebase fp8 wo quant | 3.01 |
|
29 |
+
| 1024x720 | RTX4090 | β compile blocks & extras | 3.6 |
|
30 |
| 1024x720 | RTX4090 | β
compile blocks & extras | 4.96 |
|
31 |
| 1024x720 | RTX6000ADA | bfl codebase | 2.37 |
|
32 |
+
| 1024x720 | RTX6000ADA | β compile blocks & extras | 2.87 |
|
33 |
| 1024x720 | RTX6000ADA | β
compile blocks & extras | 3.78 |
|
34 |
|
35 |
## Table of Contents
|