AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[INIT] Fix fallback for unsupported user-specified runtime unroll factor (#1780)

Browse Source 
* [INIT] Fix fallback for unsupported user-specified runtime unroll factor
* Add CollTrace guard
* Move `commSetUnrollFactor()` to rccl_wrap.cc
* Modify comments in the device-code generator script
This commit is contained in:
Nilesh M Negi
2025-07-10 10:56:18 -05:00
committed by GitHub
parent 68d6f99e0f
commit 2c099fe29a
6 changed files with 61 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/device/generate.py
+8 -9
View File
@@ -52,19 +52,21 @@ else:
# make ONLY_FUNCS="AllGather * *|Broadcast * *|SendRecv"
#
# # Only AllReduce Sum int32_t (but all algos, protos)
# make ONLY_FUNCS="AllReduce * * Sum int32_t"
# make ONLY_FUNCS="AllReduce * * Sum i32"
#
# # Only AllReduce RING Max float (but all protos and unrolls)
# make ONLY_FUNCS="AllReduce RING * Max float"
# make ONLY_FUNCS="AllReduce RING * Max f32"
#
# # AllReduce TREE LL128 Prod rccl_bfloat16 unroll=1
# make ONLY_FUNCS="AllReduce TREE LL128 Prod rccl_bfloat16 1"
# make ONLY_FUNCS="AllReduce TREE LL128 Prod bf16 1"
#
# # AllReduce RING SIMPLE and ReduceScatter RING LL float (but all redops, types, unrolls for AllReduce and all redops, unrolls for ReduceScatter)
# make ONLY_FUNCS="AllReduce RING SIMPLE * *|ReduceScatter RING LL * float *"
# make ONLY_FUNCS="AllReduce RING SIMPLE * *|ReduceScatter RING LL * f32 *"
# --- or ---
# make ONLY_FUNCS="AllReduce RING SIMPLE|ReduceScatter RING LL * float *"
# make ONLY_FUNCS="AllReduce RING/TREE LL/SIMPLE Sum/MinMax int8_t/uint8_t/half/float/double/hip_bfloat16/rccl_float8/rccl_bfloat8 1/2/4|AllGather RING LL/SIMPLE Sum int8_t 1/2/4|AllToAllPivot RING SIMPLE Sum int8_t 1/2/4|Broadcast RING LL/SIMPLE Sum int8_t 1/2/4|Reduce RING LL/SIMPLE Sum/MinMax int8_t/uint8_t/half/float/double/hip_bfloat16/rccl_float8/rccl_bfloat8 1/2/4|ReduceScatter RING LL/SIMPLE Sum/MinMax int8_t/uint8_t/half/float/double/hip_bfloat16/rccl_float8/rccl_bfloat8 1/2/4|SendRecv RING SIMPLE Sum int8_t 1/2/4"
# make ONLY_FUNCS="AllReduce RING SIMPLE|ReduceScatter RING LL * f32 *"
#
#
# make ONLY_FUNCS="AllReduce RING/TREE LL/LL128/SIMPLE Sum/MinMax i8/u8/f16/f32/f64/bf16/f8e4m3/f8e5m2 1/2/4|AllGather RING LL/LL128/SIMPLE Sum i8 1/2/4|AllToAllPivot RING SIMPLE Sum i8 1/2/4|Broadcast RING LL/LL128/SIMPLE Sum i8 1/2/4|Reduce RING LL/LL128/SIMPLE Sum/MinMax i8/u8/f16/f32/f64/bf16/f8e4m3/f8e5m2 1/2/4|ReduceScatter RING LL/LL128/SIMPLE Sum/MinMax i8/u8/f16/f32/f64/bf16/f8e4m3/f8e5m2 1/2/4|SendRecv RING SIMPLE Sum i8 1/2/4"
#
# # ONLY_FUNCS can be used together for debugging
@@ -159,9 +161,6 @@ def func_filter(function_params, current_idx, item_list=None):
# If the paramter is equal to '*', include all possible cases for it
if current_element == "*":
if current_idx == 0:
raise ValueError("Error: Paramter 'COLL' can not be type all '*'.")
# all_params list must be in the same order as function_params --> <coll> <algo> <proto> <redop> <type> <unroll>
# Get the current list from all_params
current_list = all_params[current_idx]