-
Notifications
You must be signed in to change notification settings - Fork 57
Add sliding-window attention tuning support #2439
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.
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
Open
umangyadav
wants to merge
3
commits into
develop
Choose a base branch
from
users/umayadav/sliding-window-tuning
base: develop
Could not load branches
Branch not found: {{ refName }}
Loading
Could not load tags
Nothing to show
Loading
Are you sure you want to change the base?
Some commits from the old base branch may be removed from the timeline,
and old review comments may become outdated.
Open
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit
Hold shift + click to select a range
File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
60 changes: 60 additions & 0 deletions
60
mlir/test/fusion/mixr-attention-sliding-window-problem-key.mlir
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,60 @@ | ||
| // RUN: rocmlir-driver -kernel-pipeline=migraphx,highlevel %s | rocmlir-gen --emit-tuning-key - | FileCheck %s | ||
| // CHECK: gfx942 | ||
| // CHECK-SAME: 304 | ||
| // CHECK-SAME: -t f16 -transQ false -transK true -transV false -transO false -causal false -return_lse false -split_kv 1 -sliding_window_size 3 -num_heads_q 1 -num_heads_kv 1 -g 2 -seq_len_q 1 -seq_len_k 8 -head_dim_qk 2 -head_dim_v 2 -with-attn-scale false -with-attn-bias false -transBias false | ||
| // CHECK-NOT: current_seq_len | ||
|
|
||
| module { | ||
| func.func @mlir_attention(%arg0: !migraphx.shaped<1x1xsi32, 1x1>, | ||
| %arg1: !migraphx.shaped<1x6x1x2xf16, 12x2x2x1>, | ||
| %arg2: !migraphx.shaped<1x2x8x2xf16, 32x16x2x1>, | ||
| %arg3: !migraphx.shaped<1x2x8x2xf16, 32x16x2x1>) -> !migraphx.shaped<1x1x4xf16, 4x4x1> attributes {rock.kernel, rock.arch = "gfx942", rock.num_cu = 304 : i64} { | ||
| %0 = migraphx.literal(dense<-3> : tensor<1xsi32>) : <1xsi32, 1> | ||
| %1 = migraphx.literal(dense<[0, 1, 2, 3, 4, 5, 6, 7]> : tensor<8xsi32>) : <8xsi32, 1> | ||
| %2 = migraphx.literal(dense<4> : tensor<1x1xsi32>) : <1x1xsi32, 1x1> | ||
| %3 = migraphx.literal(dense<[0, 1, 2, 3, 4, 5, 6, 7]> : tensor<8xsi32>) : <8xsi32, 1> | ||
| %4 = migraphx.literal(dense<0xFC00> : tensor<1xf16>) : <1xf16, 1> | ||
| %5 = migraphx.literal(dense<5.000000e-01> : tensor<1xf16>) : <1xf16, 1> | ||
| %6 = migraphx.multibroadcast %1 {out_dyn_dims = [], out_lens = [1, 1, 1, 8]} : <8xsi32, 1> -> <1x1x1x8xsi32, 0x0x0x1> | ||
| %7 = migraphx.multibroadcast %0 {out_dyn_dims = [], out_lens = [1, 1, 1, 1]} : <1xsi32, 1> -> <1x1x1x1xsi32, 0x0x0x1> | ||
| %8 = migraphx.reshape %arg0 {dims = [1, 1, 1, 1]} : <1x1xsi32, 1x1> -> <1x1x1x1xsi32, 1x1x1x1> | ||
| %9 = migraphx.reshape %2 {dims = [1, 1, 1, 1]} : <1x1xsi32, 1x1> -> <1x1x1x1xsi32, 1x1x1x1> | ||
| %10 = migraphx.reshape %2 {dims = [1, 1, 1, 1]} : <1x1xsi32, 1x1> -> <1x1x1x1xsi32, 1x1x1x1> | ||
| %11 = migraphx.clip %8, %9, %10 : <1x1x1x1xsi32, 1x1x1x1>, <1x1x1x1xsi32, 1x1x1x1>, <1x1x1x1xsi32, 1x1x1x1> -> <1x1x1x1xsi32, 1x1x1x1> | ||
| %12 = migraphx.slice %arg1 {axes = [1], ends = [2], starts = [0]} : <1x6x1x2xf16, 12x2x2x1> -> <1x2x1x2xf16, 12x2x2x1> | ||
| %13 = migraphx.transpose %arg2 {permutation = [0, 1, 3, 2]} : <1x2x8x2xf16, 32x16x2x1> -> <1x2x2x8xf16, 32x16x1x2> | ||
| %14 = migraphx.dot %12, %13 : <1x2x1x2xf16, 12x2x2x1>, <1x2x2x8xf16, 32x16x1x2> -> <1x2x1x8xf16, 16x8x8x1> | ||
| %15 = migraphx.multibroadcast %4 {out_dyn_dims = [], out_lens = [1, 2, 1, 8]} : <1xf16, 1> -> <1x2x1x8xf16, 0x0x0x0> | ||
| %16 = migraphx.multibroadcast %5 {out_dyn_dims = [], out_lens = [1, 2, 1, 8]} : <1xf16, 1> -> <1x2x1x8xf16, 0x0x0x0> | ||
| %17 = migraphx.mul %14, %16 : <1x2x1x8xf16, 16x8x8x1>, <1x2x1x8xf16, 0x0x0x0> -> <1x2x1x8xf16, 16x8x8x1> | ||
| %18 = migraphx.add %11, %7 : <1x1x1x1xsi32, 1x1x1x1>, <1x1x1x1xsi32, 0x0x0x1> -> <1x1x1x1xsi32, 1x1x1x1> | ||
| %19 = migraphx.multibroadcast %18 {out_dyn_dims = [], out_lens = [8, 1, 1, 1]} : <1x1x1x1xsi32, 1x1x1x1> -> <8x1x1x1xsi32, 0x1x1x1> | ||
| %20 = migraphx.reshape %19 {dims = [8]} : <8x1x1x1xsi32, 0x1x1x1> -> <8xsi32, 0> | ||
| %21 = migraphx.greater %20, %3 : <8xsi32, 0>, <8xsi32, 1> -> <8xsi32, 1> | ||
| %22 = migraphx.convert %21 {target_type = 0 : i64} : <8xsi32, 1> to <8xsi8, 1> | ||
| %23 = migraphx.broadcast %22 {axis = 3 : i64, out_lens = [1, 2, 1, 8]} : <8xsi8, 1> -> <1x2x1x8xsi8, 0x0x0x1> | ||
| %24 = migraphx.where %23, %15, %17 : <1x2x1x8xsi8, 0x0x0x1>, <1x2x1x8xf16, 0x0x0x0>, <1x2x1x8xf16, 16x8x8x1> -> <1x2x1x8xf16, 16x8x8x1> | ||
| %25 = migraphx.multibroadcast %11 {out_dyn_dims = [], out_lens = [1, 1, 1, 8]} : <1x1x1x1xsi32, 1x1x1x1> -> <1x1x1x8xsi32, 1x1x1x0> | ||
| %26 = migraphx.greater %6, %25 : <1x1x1x8xsi32, 0x0x0x1>, <1x1x1x8xsi32, 1x1x1x0> -> <1x1x1x8xsi32, 0x0x0x1> | ||
| %27 = migraphx.convert %26 {target_type = 0 : i64} : <1x1x1x8xsi32, 0x0x0x1> to <1x1x1x8xsi8, 0x0x0x1> | ||
| %28 = migraphx.multibroadcast %27 {out_dyn_dims = [], out_lens = [1, 2, 1, 8]} : <1x1x1x8xsi8, 0x0x0x1> -> <1x2x1x8xsi8, 0x0x0x1> | ||
| %29 = migraphx.where %28, %15, %24 : <1x2x1x8xsi8, 0x0x0x1>, <1x2x1x8xf16, 0x0x0x0>, <1x2x1x8xf16, 16x8x8x1> -> <1x2x1x8xf16, 16x8x8x1> | ||
| %30 = migraphx.convert %29 {target_type = 2 : i64} : <1x2x1x8xf16, 16x8x8x1> to <1x2x1x8xf32, 16x8x8x1> | ||
| %31 = migraphx.reshape %30 {dims = [1, 2, 1, 8]} : <1x2x1x8xf32, 16x8x8x1> -> <1x2x1x8xf32, 16x8x8x1> | ||
| %32 = migraphx.reduce_max %31 {axes = [3]} : <1x2x1x8xf32, 16x8x8x1> -> <1x2x1x1xf32, 2x1x1x1> | ||
| %33 = migraphx.reshape %32 {dims = [1, 2, 1, 1]} : <1x2x1x1xf32, 2x1x1x1> -> <1x2x1x1xf32, 2x1x1x1> | ||
| %34 = migraphx.multibroadcast %33 {out_dyn_dims = [], out_lens = [1, 2, 1, 8]} : <1x2x1x1xf32, 2x1x1x1> -> <1x2x1x8xf32, 2x1x1x0> | ||
| %35 = migraphx.sub %30, %34 : <1x2x1x8xf32, 16x8x8x1>, <1x2x1x8xf32, 2x1x1x0> -> <1x2x1x8xf32, 16x8x8x1> | ||
| %36 = migraphx.exp %35 : <1x2x1x8xf32, 16x8x8x1> -> <1x2x1x8xf32, 16x8x8x1> | ||
| %37 = migraphx.reshape %36 {dims = [1, 2, 1, 8]} : <1x2x1x8xf32, 16x8x8x1> -> <1x2x1x8xf32, 16x8x8x1> | ||
| %38 = migraphx.reduce_sum %37 {axes = [3]} : <1x2x1x8xf32, 16x8x8x1> -> <1x2x1x1xf32, 2x1x1x1> | ||
| %39 = migraphx.reshape %38 {dims = [1, 2, 1, 1]} : <1x2x1x1xf32, 2x1x1x1> -> <1x2x1x1xf32, 2x1x1x1> | ||
| %40 = migraphx.multibroadcast %39 {out_dyn_dims = [], out_lens = [1, 2, 1, 8]} : <1x2x1x1xf32, 2x1x1x1> -> <1x2x1x8xf32, 2x1x1x0> | ||
| %41 = migraphx.div %36, %40 : <1x2x1x8xf32, 16x8x8x1>, <1x2x1x8xf32, 2x1x1x0> -> <1x2x1x8xf32, 16x8x8x1> | ||
| %42 = migraphx.convert %41 {target_type = 1 : i64} : <1x2x1x8xf32, 16x8x8x1> to <1x2x1x8xf16, 16x8x8x1> | ||
| %43 = migraphx.dot %42, %arg3 : <1x2x1x8xf16, 16x8x8x1>, <1x2x8x2xf16, 32x16x2x1> -> <1x2x1x2xf16, 4x2x2x1> | ||
| %44 = migraphx.transpose %43 {permutation = [0, 2, 1, 3]} : <1x2x1x2xf16, 4x2x2x1> -> <1x1x2x2xf16, 4x2x2x1> | ||
| %45 = migraphx.reshape %44 {dims = [1, 1, 4]} : <1x1x2x2xf16, 4x2x2x1> -> <1x1x4xf16, 4x4x1> | ||
| return %45 : !migraphx.shaped<1x1x4xf16, 4x4x1> | ||
| } | ||
| } |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -5,9 +5,10 @@ | |
| # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||
| """Pure-Python coverage for attention tuning DB compatibility. | ||
|
|
||
| The tuning DB key for attention has grown optional boolean flags over time | ||
| (``-with-attn-scale``, ``-with-attn-bias``, ``-transBias``). False-valued flags | ||
| are identity cases, so old DB rows that omit them should still be readable: | ||
| The tuning DB key for attention has grown optional fields over time | ||
| (``-with-attn-scale``, ``-with-attn-bias``, ``-transBias``, and | ||
| ``-sliding_window_size``). Disabled fields are identity cases, so old DB rows | ||
| that omit them should still be readable: | ||
| ``read_tuning_db`` canonicalizes every stored key through | ||
| ``AttentionConfiguration``, which re-adds the (default false) flags. True-valued | ||
| flags describe different generated kernels and must not be silently matched | ||
|
|
@@ -87,10 +88,10 @@ def setUp(self): | |
| """Give each test its own temporary file prefix.""" | ||
| self.tmp_prefix = Path(f"{TMP_PREFIX}.{self._testMethodName}") | ||
|
|
||
| def read_db_from_legacy_key(self, legacy_key): | ||
| """Read a tuning DB written with a legacy key (canonicalized on read).""" | ||
| def read_db_from_key(self, key): | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Why rename this function? Is it being used in non-legacy cases as well?
Member
Author
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Yes It was renamed because it now handles both legacy and current keys. |
||
| """Read a tuning DB whose key is canonicalized on read.""" | ||
| path = Path(f"{self.tmp_prefix}.tsv") | ||
| write_tuning_db(path, legacy_key) | ||
| write_tuning_db(path, key) | ||
| return read_tuning_db(str(path), AttentionConfiguration, NUM_CU, NUM_CHIPLETS) | ||
|
|
||
| def db_key(self, config): | ||
|
|
@@ -103,15 +104,15 @@ def test_read_tuning_db_matches_legacy_all_false_attention_flags(self): | |
| legacy_key = drop_flags(current_config.to_command_line(), " -with-attn-scale false", | ||
| " -with-attn-bias false", " -transBias false") | ||
|
|
||
| db = self.read_db_from_legacy_key(legacy_key) | ||
| db = self.read_db_from_key(legacy_key) | ||
| self.assertEqual(db.get(self.db_key(current_config)), PERFCONFIG) | ||
|
|
||
| def test_read_tuning_db_matches_pre_transbias_scale_bias_key(self): | ||
| """Scale+bias DB rows from before transBias still match non-transposed bias.""" | ||
| current_config = make_config("-with-attn-scale true -with-attn-bias true -transBias false") | ||
| legacy_key = drop_flags(current_config.to_command_line(), " -transBias false") | ||
|
|
||
| db = self.read_db_from_legacy_key(legacy_key) | ||
| db = self.read_db_from_key(legacy_key) | ||
| self.assertEqual(db.get(self.db_key(current_config)), PERFCONFIG) | ||
|
|
||
| def test_read_tuning_db_keeps_true_scale_bias_distinct(self): | ||
|
|
@@ -125,7 +126,7 @@ def test_read_tuning_db_keeps_true_scale_bias_distinct(self): | |
| scale_bias_config = make_config( | ||
| "-with-attn-scale true -with-attn-bias true -transBias false") | ||
|
|
||
| db = self.read_db_from_legacy_key(legacy_all_false_key) | ||
| db = self.read_db_from_key(legacy_all_false_key) | ||
| self.assertNotIn(self.db_key(scale_bias_config), db) | ||
|
|
||
| def test_read_tuning_db_keeps_true_trans_bias_distinct(self): | ||
|
|
@@ -140,11 +141,46 @@ def test_read_tuning_db_keeps_true_trans_bias_distinct(self): | |
| "-with-attn-scale false -with-attn-bias true -transBias true") | ||
|
|
||
| self.assertIn("-transBias true", trans_bias_config.to_command_line()) | ||
| db = self.read_db_from_legacy_key(legacy_all_false_key) | ||
| db = self.read_db_from_key(legacy_all_false_key) | ||
| self.assertNotIn(self.db_key(trans_bias_config), db) | ||
|
|
||
| def test_quick_tuning_gen_defaults_missing_trans_bias_column(self): | ||
| """Legacy debug TSV rows without TransBias get a false default.""" | ||
| def test_read_tuning_db_keeps_sliding_window_distinct(self): | ||
| """A sliding-window kernel must not match a row without its window.""" | ||
| no_window_config = make_config( | ||
| "-with-attn-scale false -with-attn-bias false -transBias false") | ||
| sliding_window_config = make_config("-sliding_window_size 8 -with-attn-scale false " | ||
| "-with-attn-bias false -transBias false") | ||
|
|
||
| db = self.read_db_from_key(no_window_config.to_command_line()) | ||
| self.assertNotIn(self.db_key(sliding_window_config), db) | ||
|
|
||
| def test_read_tuning_db_matches_pre_transbias_sliding_window_key(self): | ||
| """A sliding-window row from before transBias must still match.""" | ||
| current_config = make_config("-sliding_window_size 8 -with-attn-scale false " | ||
| "-with-attn-bias false -transBias false") | ||
| legacy_key = drop_flags(current_config.to_command_line(), " -transBias false") | ||
|
|
||
| self.assertIn("-sliding_window_size 8", legacy_key) | ||
| db = self.read_db_from_key(legacy_key) | ||
| self.assertEqual(db.get(self.db_key(current_config)), PERFCONFIG) | ||
|
|
||
| def test_current_seq_len_is_runtime_only(self): | ||
| """Runtime positions reach rocmlir-gen without entering the tuning key.""" | ||
| config = make_config("-sliding_window_size 8 -current_seq_len 4 " | ||
| "-with-attn-scale false -with-attn-bias false -transBias false") | ||
|
|
||
| key = config.to_command_line() | ||
| driver_args = config.generate_mlir_driver_commandline("", kernel_repeats=None).split() | ||
| report_entry = config.table_entry(1.0) | ||
| self.assertIn("-sliding_window_size 8", key) | ||
| self.assertNotIn("current_seq_len", key) | ||
| self.assertIn("-sliding_window_size=8", driver_args) | ||
| self.assertEqual(driver_args.count("-current_seq_len=4"), 1) | ||
| self.assertEqual(report_entry["SlidingWindowSize"], 8) | ||
| self.assertNotIn("CurrentSeqLen", report_entry) | ||
|
|
||
| def test_quick_tuning_gen_defaults_missing_optional_columns(self): | ||
| """Legacy debug TSV rows get disabled optional-field defaults.""" | ||
| debug_path = Path(f"{self.tmp_prefix}.debug") | ||
| debug_path.write_text( | ||
| "DataType\tChip\tnumCU\tnumChiplets\tTransQ\tTransK\tTransV\tTransO\t" | ||
|
|
@@ -154,17 +190,15 @@ def test_quick_tuning_gen_defaults_missing_trans_bias_column(self): | |
| f"False\tFalse\t1\tTrue\tTrue\t1\t16\t16\t1\t1\t32\t32\t{PERFCONFIG}\t1.0\n") | ||
|
|
||
| df = load_data([str(debug_path)], no_splitk=False) | ||
| self.assertIn("TransBias", df.columns) | ||
| self.assertTrue(df["TransBias"].eq(False).all()) | ||
| self.assertTrue(df["SlidingWindowSize"].eq(0).all()) | ||
|
|
||
| grouped = df.groupby(get_target_columns("attention") + ["PerfConfig"], | ||
| as_index=False)["TFlops"].max() | ||
| self.assertFalse(grouped.empty) | ||
|
|
||
| def test_quick_tuning_gen_fills_trans_bias_nan_in_mixed_files(self): | ||
| """Mixing legacy (no TransBias) and new (TransBias) TSVs must not drop | ||
| the legacy rows: concat leaves NaN in the legacy rows, and TransBias is | ||
| a groupby key, so an unfilled NaN would be silently dropped.""" | ||
| def test_quick_tuning_gen_fills_optional_nan_in_mixed_files(self): | ||
| """Mixed legacy/current TSVs must retain rows with missing fields.""" | ||
| cols_no_tb = ("DataType\tChip\tnumCU\tnumChiplets\tTransQ\tTransK\tTransV\tTransO\t" | ||
| "Causal\tReturnLSE\tSplitKV\tWithAttnScale\tWithAttnBias\tG\tSeqLenQ\t" | ||
| "SeqLenK\tNumHeadsQ\tNumHeadsKV\tHeadDimQK\tHeadDimV\tPerfConfig\tTFlops\n") | ||
|
|
@@ -173,23 +207,25 @@ def test_quick_tuning_gen_fills_trans_bias_nan_in_mixed_files(self): | |
| cols_no_tb + f"f16\tgfx950\t{NUM_CU}\t{NUM_CHIPLETS}\tFalse\tFalse\tFalse\tFalse\t" | ||
| f"False\tFalse\t1\tTrue\tTrue\t1\t16\t16\t1\t1\t32\t32\t{PERFCONFIG}\t1.0\n") | ||
|
|
||
| cols_tb = ( | ||
| cols_optional = ( | ||
| "DataType\tChip\tnumCU\tnumChiplets\tTransQ\tTransK\tTransV\tTransO\t" | ||
| "Causal\tReturnLSE\tSplitKV\tWithAttnScale\tWithAttnBias\tTransBias\tG\tSeqLenQ\t" | ||
| "SeqLenK\tNumHeadsQ\tNumHeadsKV\tHeadDimQK\tHeadDimV\tPerfConfig\tTFlops\n") | ||
| "Causal\tReturnLSE\tSplitKV\tSlidingWindowSize\tWithAttnScale\tWithAttnBias\t" | ||
| "TransBias\tG\tSeqLenQ\tSeqLenK\tNumHeadsQ\tNumHeadsKV\tHeadDimQK\tHeadDimV\t" | ||
| "PerfConfig\tTFlops\n") | ||
| new_path = Path(f"{self.tmp_prefix}.new.debug") | ||
| new_path.write_text( | ||
| cols_tb + f"f16\tgfx950\t{NUM_CU}\t{NUM_CHIPLETS}\tFalse\tFalse\tFalse\tFalse\t" | ||
| f"False\tFalse\t1\tTrue\tTrue\tTrue\t1\t16\t16\t1\t1\t32\t32\t{PERFCONFIG}\t2.0\n") | ||
| new_path.write_text(cols_optional + | ||
| f"f16\tgfx950\t{NUM_CU}\t{NUM_CHIPLETS}\tFalse\tFalse\tFalse\tFalse\t" | ||
| f"False\tFalse\t1\t8\tTrue\tTrue\tTrue\t1\t16\t16\t1\t1\t32\t32\t" | ||
| f"{PERFCONFIG}\t2.0\n") | ||
|
|
||
| df = load_data([str(legacy_path), str(new_path)], no_splitk=False) | ||
| self.assertFalse(df["TransBias"].isna().any()) | ||
| self.assertFalse(df["SlidingWindowSize"].isna().any()) | ||
|
|
||
| grouped = df.groupby(get_target_columns("attention") + ["PerfConfig"], | ||
| as_index=False)["TFlops"].max() | ||
| # Both problems survive: legacy TransBias=False and new TransBias=True. | ||
| # Without the NaN fill, the legacy row would be dropped and only one | ||
| # group would remain. | ||
| # Both problems survive: legacy fields default to disabled, while the | ||
| # current row has TransBias=True and SlidingWindowSize=8. | ||
| self.assertEqual(len(grouped), 2) | ||
|
|
||
|
|
||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Nit: I think that we can drop this comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Removed.