Add tuned FP8 config for HF RTX PRO 6000 Server Edition
Browse files
serving/configs/N=12576,K=4096,device_name=NVIDIA_RTX_PRO_6000_Blackwell_Server_Edition,dtype=fp8_w8a8,block_shape=[32,32].json
ADDED
|
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"1": {
|
| 3 |
+
"BLOCK_SIZE_M": 16,
|
| 4 |
+
"BLOCK_SIZE_N": 64,
|
| 5 |
+
"BLOCK_SIZE_K": 32,
|
| 6 |
+
"GROUP_SIZE_M": 1,
|
| 7 |
+
"num_warps": 4,
|
| 8 |
+
"num_stages": 6
|
| 9 |
+
},
|
| 10 |
+
"2": {
|
| 11 |
+
"BLOCK_SIZE_M": 16,
|
| 12 |
+
"BLOCK_SIZE_N": 128,
|
| 13 |
+
"BLOCK_SIZE_K": 32,
|
| 14 |
+
"GROUP_SIZE_M": 32,
|
| 15 |
+
"num_warps": 4,
|
| 16 |
+
"num_stages": 4
|
| 17 |
+
},
|
| 18 |
+
"4": {
|
| 19 |
+
"BLOCK_SIZE_M": 16,
|
| 20 |
+
"BLOCK_SIZE_N": 128,
|
| 21 |
+
"BLOCK_SIZE_K": 32,
|
| 22 |
+
"GROUP_SIZE_M": 32,
|
| 23 |
+
"num_warps": 4,
|
| 24 |
+
"num_stages": 4
|
| 25 |
+
},
|
| 26 |
+
"8": {
|
| 27 |
+
"BLOCK_SIZE_M": 16,
|
| 28 |
+
"BLOCK_SIZE_N": 128,
|
| 29 |
+
"BLOCK_SIZE_K": 32,
|
| 30 |
+
"GROUP_SIZE_M": 1,
|
| 31 |
+
"num_warps": 4,
|
| 32 |
+
"num_stages": 4
|
| 33 |
+
},
|
| 34 |
+
"16": {
|
| 35 |
+
"BLOCK_SIZE_M": 16,
|
| 36 |
+
"BLOCK_SIZE_N": 128,
|
| 37 |
+
"BLOCK_SIZE_K": 32,
|
| 38 |
+
"GROUP_SIZE_M": 1,
|
| 39 |
+
"num_warps": 4,
|
| 40 |
+
"num_stages": 6
|
| 41 |
+
},
|
| 42 |
+
"64": {
|
| 43 |
+
"BLOCK_SIZE_M": 64,
|
| 44 |
+
"BLOCK_SIZE_N": 64,
|
| 45 |
+
"BLOCK_SIZE_K": 32,
|
| 46 |
+
"GROUP_SIZE_M": 8,
|
| 47 |
+
"num_warps": 4,
|
| 48 |
+
"num_stages": 3
|
| 49 |
+
},
|
| 50 |
+
"256": {
|
| 51 |
+
"BLOCK_SIZE_M": 128,
|
| 52 |
+
"BLOCK_SIZE_N": 64,
|
| 53 |
+
"BLOCK_SIZE_K": 32,
|
| 54 |
+
"GROUP_SIZE_M": 32,
|
| 55 |
+
"num_warps": 4,
|
| 56 |
+
"num_stages": 2
|
| 57 |
+
},
|
| 58 |
+
"1024": {
|
| 59 |
+
"BLOCK_SIZE_M": 64,
|
| 60 |
+
"BLOCK_SIZE_N": 128,
|
| 61 |
+
"BLOCK_SIZE_K": 32,
|
| 62 |
+
"GROUP_SIZE_M": 8,
|
| 63 |
+
"num_warps": 4,
|
| 64 |
+
"num_stages": 2
|
| 65 |
+
},
|
| 66 |
+
"2048": {
|
| 67 |
+
"BLOCK_SIZE_M": 128,
|
| 68 |
+
"BLOCK_SIZE_N": 128,
|
| 69 |
+
"BLOCK_SIZE_K": 32,
|
| 70 |
+
"GROUP_SIZE_M": 8,
|
| 71 |
+
"num_warps": 4,
|
| 72 |
+
"num_stages": 2
|
| 73 |
+
},
|
| 74 |
+
"4096": {
|
| 75 |
+
"BLOCK_SIZE_M": 128,
|
| 76 |
+
"BLOCK_SIZE_N": 128,
|
| 77 |
+
"BLOCK_SIZE_K": 32,
|
| 78 |
+
"GROUP_SIZE_M": 8,
|
| 79 |
+
"num_warps": 4,
|
| 80 |
+
"num_stages": 2
|
| 81 |
+
},
|
| 82 |
+
"8192": {
|
| 83 |
+
"BLOCK_SIZE_M": 128,
|
| 84 |
+
"BLOCK_SIZE_N": 128,
|
| 85 |
+
"BLOCK_SIZE_K": 32,
|
| 86 |
+
"GROUP_SIZE_M": 1,
|
| 87 |
+
"num_warps": 4,
|
| 88 |
+
"num_stages": 2
|
| 89 |
+
}
|
| 90 |
+
}
|