Skip to content

Commit

Permalink
Simple num_stages fix without re-tuning for performance
Browse files Browse the repository at this point in the history
  • Loading branch information
@gshtras
gshtras committed Dec 19, 2024
1 parent ae77a01 commit c1ccc5e
Show file tree
Hide file tree
Showing 43 changed files with 789 additions and 789 deletions.
36 changes: 18 additions & 18 deletions .../model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 1
Expand All @@ -16,7 +16,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -27,7 +27,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -38,7 +38,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -49,7 +49,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -60,7 +60,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -71,7 +71,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 2,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 1
Expand All @@ -82,7 +82,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 2,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -93,7 +93,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -104,7 +104,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -115,7 +115,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -126,7 +126,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 1
Expand All @@ -137,7 +137,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -148,7 +148,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 32,
"kpack": 2
Expand All @@ -159,7 +159,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -170,7 +170,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 2
Expand All @@ -181,7 +181,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 1
Expand All @@ -192,7 +192,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 16,
"kpack": 1
Expand Down
36 changes: 18 additions & 18 deletions ...ers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X_OAM,dtype=fp8_w8a8.json
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"2": {
Expand All @@ -14,7 +14,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"4": {
Expand All @@ -23,7 +23,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"8": {
Expand All @@ -32,7 +32,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"16": {
Expand All @@ -41,7 +41,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"24": {
Expand All @@ -50,7 +50,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"32": {
Expand All @@ -59,7 +59,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"48": {
Expand All @@ -68,7 +68,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"64": {
Expand All @@ -77,7 +77,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"num_warps": 1,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"96": {
Expand All @@ -86,7 +86,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"num_warps": 2,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"128": {
Expand All @@ -95,7 +95,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"256": {
Expand All @@ -104,7 +104,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"512": {
Expand All @@ -113,7 +113,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"1024": {
Expand All @@ -122,7 +122,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"1536": {
Expand All @@ -131,7 +131,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"2048": {
Expand All @@ -140,7 +140,7 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"3072": {
Expand All @@ -149,7 +149,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
},
"4096": {
Expand All @@ -158,7 +158,7 @@
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 0,
"num_stages": 2,
"waves_per_eu": 0
}
}
Loading

0 comments on commit c1ccc5e

Please sign in to comment.