[wip] make Float8Linear amax init more FSDP+compile friendly

float8_experimental/float8_utils.py

@@ -7,6 +7,7 @@
 
 import torch
 import torch.distributed as dist
+import torch.distributed._functional_collectives as _functional_collectives
 
 # Helpful visualizer for debugging (only supports fp32):
 # https://www.h-schmidt.net/FloatConverter/IEEE754.html
@@ -60,7 +61,12 @@ def tensor_to_amax(x, distributed_reduction=False):
     # If the user did not ask for it, assume that it will
     # happen elsewhere.
     if distributed_reduction and dist.is_initialized():
-        dist.all_reduce(amax, op=dist.ReduceOp.MAX)
+        # TODO(future): support process groups
+        ranks = list(range(dist.get_world_size()))
+        # print('ranks', ranks)
+        # print('old amax', amax)
+        amax = _functional_collectives.all_reduce(amax, "max", group=ranks)
+        # print('new amax', amax)
 
     return amax
 

test/test_fsdp.py

@@ -104,10 +104,13 @@ def fsdp_main(rank, world_size, args):
     ref_input_local = ref_input_global[bsz_local_start:bsz_local_end].to(rank)
 
     sync_float8_func = sync_float8_amax_and_scale_history
-    if compile:
+    if compile and False:
         sync_float8_func = torch.compile(
             sync_float8_amax_and_scale_history, fullgraph=fullgraph
         )
+        model = torch.compile(model)
+
+    print(model)
 
     def forward_backward(model):
         optimizer.zero_grad()
@@ -118,14 +121,19 @@ def forward_backward(model):
         return y_local
 
     for iter in range(N_ITER):
-        # We first run one iteration without compile, as a workaround to compile float8 layer.
-        # In the first iter, float8 layers go to the branches of "self.is_amax_initialized == False"
-        # After that, float8 layers go the the branches of "self.is_amax_initialized == True"
-        # TODO: Need to fix compile to run wihtout this workaround.
-        if iter == 1 and compile:
-            model = torch.compile(model, fullgraph=fullgraph)
         y_local = forward_backward(model)
 
+    if compile and False:
+        base = model._orig_mod._fsdp_wrapped_module
+    else:
+        base = model._fsdp_wrapped_module
+    print('0_x', base[0].fp8_amax_history_x)
+    print('2_x', base[2].fp8_amax_history_x)
+    print('0_w', base[0].fp8_amax_history_w)
+    print('2_w', base[2].fp8_amax_history_w)
+    print('0_g', base[0].fp8_amax_history_dL_dY)
+    print('2_g', base[2].fp8_amax_history_dL_dY)
+
     # get global y
     y_global = [
         torch.zeros(*y_local.shape, dtype=base_dtype).to(rank)

test/test_fsdp.sh

@@ -7,12 +7,12 @@ launch() {
     echo "launching IS_FP8 $IS_FP8, compile_fsdp $COMPILE, fullgraph $FULLGRAPH"
 
     # generate the test data
-    python test/test_fsdp.py --mode generate --is_fp8 $IS_FP8 --compile_fsdp $COMPILE --fullgraph $FULLGRAPH
-    echo "Success: ✅"
+    # python test/test_fsdp.py --mode generate --is_fp8 $IS_FP8 --compile_fsdp $COMPILE --fullgraph $FULLGRAPH
+    # echo "Success: ✅"
 
     # generate single GPU model output and updated state dict
-    python test/test_fsdp.py --mode single_gpu --is_fp8 $IS_FP8 --compile_fsdp $COMPILE --fullgraph $FULLGRAPH
-    echo "Success: ✅"
+    # python test/test_fsdp.py --mode single_gpu --is_fp8 $IS_FP8 --compile_fsdp $COMPILE --fullgraph $FULLGRAPH
+    # echo "Success: ✅"
 
     # generate FSDP model output and updated state dict
     # the NCCL_DEBUG setting is to avoid log spew
@@ -30,7 +30,8 @@ launch() {
 }
 
 # IS_FP8, COMPILE, FULLGRAPH
-for i in False,False,False True,False,False True,True,False
+# for i in False,False,False True,False,False True,True,False
+for i in True,True,False
 do
     IFS=","; set -- $i;
     IS_FP8=$1; COMPILE=$2; FULLGRAPH=$3