Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add llama and resnet explorer tests #1286

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Add llama and resnet explorer tests #1286

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
17d58da
Add llama and resnet explorer tests
odjuricicTT Nov 15, 2024
71aa1b7
Remove system desc from tests
odjuricicTT Nov 18, 2024
20465e0
Merge branch 'main' into odjuricic/explorer-models
odjuricicTT Nov 19, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
120 changes: 120 additions & 0 deletions tools/explorer/test/models/llama_attention_no_rot_emb_ttir.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,120 @@
#any_device = #tt.operand_constraint<dram|l1|scalar|tile|none|interleaved|single_bank|height_sharded|width_sharded|block_sharded|any_layout|any_device|any_device_tile|l1_block_sharded>
#loc = loc("SelfAttention":0:0)
module @SelfAttention attributes {} {
func.func @forward(%arg0: tensor<1x12x3200xf32> {ttir.name = "hidden_states_1"} loc("SelfAttention":0:0), %arg1: tensor<1x1x12x12xf32> {ttir.name = "attention_mask"} loc("SelfAttention":0:0), %arg2: tensor<1xf32> {ttir.name = "input_1_multiply_20"} loc("SelfAttention":0:0), %arg3: tensor<3200x3200xf32> {ttir.name = "model.q_proj.weight"} loc("SelfAttention":0:0), %arg4: tensor<3200x3200xf32> {ttir.name = "model.k_proj.weight"} loc("SelfAttention":0:0), %arg5: tensor<3200x3200xf32> {ttir.name = "model.v_proj.weight"} loc("SelfAttention":0:0), %arg6: tensor<3200x3200xf32> {ttir.name = "model.o_proj.weight"} loc("SelfAttention":0:0)) -> (tensor<1x12x3200xf32> {ttir.name = "SelfAttention.output_reshape_38"}) {
%0 = tensor.empty() : tensor<12x3200xf32> loc(#loc30)
%1 = "ttir.squeeze"(%arg0, %0) <{dim = 0 : si32, operand_constraints = [#any_device, #any_device, #any_device, #any_device]}> : (tensor<1x12x3200xf32>, tensor<12x3200xf32>) -> tensor<12x3200xf32> loc(#loc30)
%2 = tensor.empty() : tensor<12x3200xf32> loc(#loc31)
%3 = "ttir.matmul"(%1, %arg3, %2) <{operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<12x3200xf32>, tensor<3200x3200xf32>, tensor<12x3200xf32>) -> tensor<12x3200xf32> loc(#loc31)
%4 = tensor.empty() : tensor<1x12x32x100xf32> loc(#loc32)
%5 = "ttir.reshape"(%3, %4) <{operand_constraints = [#any_device, #any_device], shape = [1 : i32, 12 : i32, 32 : i32, 100 : i32]}> : (tensor<12x3200xf32>, tensor<1x12x32x100xf32>) -> tensor<1x12x32x100xf32> loc(#loc32)
%6 = tensor.empty() : tensor<1x32x12x100xf32> loc(#loc33)
%7 = "ttir.transpose"(%5, %6) <{dim0 = -3 : si32, dim1 = -2 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x12x32x100xf32>, tensor<1x32x12x100xf32>) -> tensor<1x32x12x100xf32> loc(#loc33)
%8 = tensor.empty() : tensor<32x12x100xf32> loc(#loc34)
%9 = "ttir.squeeze"(%7, %8) <{dim = 0 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x32x12x100xf32>, tensor<32x12x100xf32>) -> tensor<32x12x100xf32> loc(#loc34)
%10 = tensor.empty() : tensor<12x3200xf32> loc(#loc35)
%11 = "ttir.matmul"(%1, %arg4, %10) <{operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<12x3200xf32>, tensor<3200x3200xf32>, tensor<12x3200xf32>) -> tensor<12x3200xf32> loc(#loc35)
%12 = tensor.empty() : tensor<1x12x32x100xf32> loc(#loc36)
%13 = "ttir.reshape"(%11, %12) <{operand_constraints = [#any_device, #any_device], shape = [1 : i32, 12 : i32, 32 : i32, 100 : i32]}> : (tensor<12x3200xf32>, tensor<1x12x32x100xf32>) -> tensor<1x12x32x100xf32> loc(#loc36)
%14 = tensor.empty() : tensor<1x32x12x100xf32> loc(#loc37)
%15 = "ttir.transpose"(%13, %14) <{dim0 = -3 : si32, dim1 = -2 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x12x32x100xf32>, tensor<1x32x12x100xf32>) -> tensor<1x32x12x100xf32> loc(#loc37)
%16 = tensor.empty() : tensor<32x12x100xf32> loc(#loc38)
%17 = "ttir.squeeze"(%15, %16) <{dim = 0 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x32x12x100xf32>, tensor<32x12x100xf32>) -> tensor<32x12x100xf32> loc(#loc38)
%18 = tensor.empty() : tensor<32x100x12xf32> loc(#loc39)
%19 = "ttir.transpose"(%17, %18) <{dim0 = -2 : si32, dim1 = -1 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<32x12x100xf32>, tensor<32x100x12xf32>) -> tensor<32x100x12xf32> loc(#loc39)
%20 = tensor.empty() : tensor<32x12x12xf32> loc(#loc40)
%21 = "ttir.matmul"(%9, %19, %20) <{operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<32x12x100xf32>, tensor<32x100x12xf32>, tensor<32x12x12xf32>) -> tensor<32x12x12xf32> loc(#loc40)
%22 = tensor.empty() : tensor<1x32x12x12xf32> loc(#loc41)
%23 = "ttir.unsqueeze"(%21, %22) <{dim = 0 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<32x12x12xf32>, tensor<1x32x12x12xf32>) -> tensor<1x32x12x12xf32> loc(#loc41)
%24 = tensor.empty() : tensor<1x32x12x12xf32> loc(#loc42)
%25 = "ttir.multiply"(%23, %arg2, %24) <{operandSegmentSizes = array<i32: 2, 1>, operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<1x32x12x12xf32>, tensor<1xf32>, tensor<1x32x12x12xf32>) -> tensor<1x32x12x12xf32> loc(#loc42)
%26 = tensor.empty() : tensor<1x32x12x12xf32> loc(#loc43)
%27 = "ttir.add"(%25, %arg1, %26) <{operandSegmentSizes = array<i32: 2, 1>, operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<1x32x12x12xf32>, tensor<1x1x12x12xf32>, tensor<1x32x12x12xf32>) -> tensor<1x32x12x12xf32> loc(#loc43)
%28 = tensor.empty() : tensor<1x32x12x12xf32> loc(#loc44)
%29 = "ttir.softmax"(%27, %28) <{dimension = -1 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x32x12x12xf32>, tensor<1x32x12x12xf32>) -> tensor<1x32x12x12xf32> loc(#loc44)
%30 = tensor.empty() : tensor<32x12x12xf32> loc(#loc45)
%31 = "ttir.squeeze"(%29, %30) <{dim = 0 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x32x12x12xf32>, tensor<32x12x12xf32>) -> tensor<32x12x12xf32> loc(#loc45)
%32 = tensor.empty() : tensor<12x3200xf32> loc(#loc46)
%33 = "ttir.matmul"(%1, %arg5, %32) <{operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<12x3200xf32>, tensor<3200x3200xf32>, tensor<12x3200xf32>) -> tensor<12x3200xf32> loc(#loc46)
%34 = tensor.empty() : tensor<1x12x32x100xf32> loc(#loc47)
%35 = "ttir.reshape"(%33, %34) <{operand_constraints = [#any_device, #any_device], shape = [1 : i32, 12 : i32, 32 : i32, 100 : i32]}> : (tensor<12x3200xf32>, tensor<1x12x32x100xf32>) -> tensor<1x12x32x100xf32> loc(#loc47)
%36 = tensor.empty() : tensor<1x32x12x100xf32> loc(#loc48)
%37 = "ttir.transpose"(%35, %36) <{dim0 = -3 : si32, dim1 = -2 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x12x32x100xf32>, tensor<1x32x12x100xf32>) -> tensor<1x32x12x100xf32> loc(#loc48)
%38 = tensor.empty() : tensor<1x32x100x12xf32> loc(#loc49)
%39 = "ttir.transpose"(%37, %38) <{dim0 = -2 : si32, dim1 = -1 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x32x12x100xf32>, tensor<1x32x100x12xf32>) -> tensor<1x32x100x12xf32> loc(#loc49)
%40 = tensor.empty() : tensor<32x100x12xf32> loc(#loc50)
%41 = "ttir.squeeze"(%39, %40) <{dim = 0 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x32x100x12xf32>, tensor<32x100x12xf32>) -> tensor<32x100x12xf32> loc(#loc50)
%42 = tensor.empty() : tensor<32x12x100xf32> loc(#loc51)
%43 = "ttir.transpose"(%41, %42) <{dim0 = -2 : si32, dim1 = -1 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<32x100x12xf32>, tensor<32x12x100xf32>) -> tensor<32x12x100xf32> loc(#loc51)
%44 = tensor.empty() : tensor<32x12x100xf32> loc(#loc52)
%45 = "ttir.matmul"(%31, %43, %44) <{operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<32x12x12xf32>, tensor<32x12x100xf32>, tensor<32x12x100xf32>) -> tensor<32x12x100xf32> loc(#loc52)
%46 = tensor.empty() : tensor<1x32x12x100xf32> loc(#loc53)
%47 = "ttir.unsqueeze"(%45, %46) <{dim = 0 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<32x12x100xf32>, tensor<1x32x12x100xf32>) -> tensor<1x32x12x100xf32> loc(#loc53)
%48 = tensor.empty() : tensor<1x12x32x100xf32> loc(#loc54)
%49 = "ttir.transpose"(%47, %48) <{dim0 = -3 : si32, dim1 = -2 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<1x32x12x100xf32>, tensor<1x12x32x100xf32>) -> tensor<1x12x32x100xf32> loc(#loc54)
%50 = tensor.empty() : tensor<12x3200xf32> loc(#loc55)
%51 = "ttir.reshape"(%49, %50) <{operand_constraints = [#any_device, #any_device], shape = [12 : i32, 3200 : i32]}> : (tensor<1x12x32x100xf32>, tensor<12x3200xf32>) -> tensor<12x3200xf32> loc(#loc55)
%52 = tensor.empty() : tensor<12x3200xf32> loc(#loc56)
%53 = "ttir.matmul"(%51, %arg6, %52) <{operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<12x3200xf32>, tensor<3200x3200xf32>, tensor<12x3200xf32>) -> tensor<12x3200xf32> loc(#loc56)
%54 = tensor.empty() : tensor<1x12x3200xf32> loc(#loc57)
%55 = "ttir.unsqueeze"(%53, %54) <{dim = 0 : si32, operand_constraints = [#any_device, #any_device]}> : (tensor<12x3200xf32>, tensor<1x12x3200xf32>) -> tensor<1x12x3200xf32> loc(#loc57)
return %55 : tensor<1x12x3200xf32> loc(#loc29)
} loc(#loc)
} loc(#loc)
#loc1 = loc("forward":4294967295:63)
#loc2 = loc("forward":4294967295:65)
#loc3 = loc("forward":4294967295:66)
#loc4 = loc("forward":4294967295:67)
#loc5 = loc("forward":4294967295:68)
#loc6 = loc("forward":4294967295:70)
#loc7 = loc("forward":4294967295:71)
#loc8 = loc("forward":4294967295:72)
#loc9 = loc("forward":4294967295:73)
#loc10 = loc("forward":4294967295:74)
#loc11 = loc("forward":4294967295:75)
#loc12 = loc("forward":4294967295:76)
#loc13 = loc("forward":4294967295:78)
#loc14 = loc("forward":4294967295:80)
#loc15 = loc("forward":4294967295:81)
#loc16 = loc("forward":4294967295:82)
#loc17 = loc("forward":4294967295:84)
#loc18 = loc("forward":4294967295:85)
#loc19 = loc("forward":4294967295:86)
#loc20 = loc("forward":4294967295:87)
#loc21 = loc("forward":4294967295:88)
#loc22 = loc("forward":4294967295:89)
#loc23 = loc("forward":4294967295:90)
#loc24 = loc("forward":4294967295:91)
#loc25 = loc("forward":4294967295:92)
#loc26 = loc("forward":4294967295:93)
#loc27 = loc("forward":4294967295:95)
#loc28 = loc("forward":4294967295:96)
#loc29 = loc(unknown)
#loc30 = loc("reshape_6.dc.squeeze.0"(#loc1))
#loc31 = loc("matmul_8"(#loc2))
#loc32 = loc("reshape_9"(#loc3))
#loc33 = loc("transpose_10"(#loc4))
#loc34 = loc("reshape_11.dc.squeeze.0"(#loc5))
#loc35 = loc("matmul_13"(#loc6))
#loc36 = loc("reshape_14"(#loc7))
#loc37 = loc("transpose_15"(#loc8))
#loc38 = loc("reshape_16.dc.squeeze.0"(#loc9))
#loc39 = loc("transpose_17"(#loc10))
#loc40 = loc("matmul_18"(#loc11))
#loc41 = loc("reshape_19.dc.unsqueeze.0"(#loc12))
#loc42 = loc("multiply_20"(#loc13))
#loc43 = loc("add_21"(#loc14))
#loc44 = loc("softmax_22"(#loc15))
#loc45 = loc("reshape_24.dc.squeeze.0"(#loc16))
#loc46 = loc("matmul_26"(#loc17))
#loc47 = loc("reshape_27"(#loc18))
#loc48 = loc("transpose_28"(#loc19))
#loc49 = loc("transpose_29"(#loc20))
#loc50 = loc("reshape_30.dc.squeeze.0"(#loc21))
#loc51 = loc("transpose_31"(#loc22))
#loc52 = loc("matmul_32"(#loc23))
#loc53 = loc("reshape_33.dc.unsqueeze.0"(#loc24))
#loc54 = loc("transpose_34"(#loc25))
#loc55 = loc("reshape_35"(#loc26))
#loc56 = loc("matmul_37"(#loc27))
#loc57 = loc("reshape_38.dc.unsqueeze.0"(#loc28))
3 changes: 1 addition & 2 deletions tools/explorer/test/models/open_llama_3b_single_layer.mlir
View file
Open in desktop

Large diffs are not rendered by default.

8,364 changes: 8,364 additions & 0 deletions tools/explorer/test/models/open_llama_3b_ttir.mlir
View file
Open in desktop

Large diffs are not rendered by default.

Loading
Loading