internal pipelining more similar to micro-benchmarks

apex/contrib/optimizers/distributed_fused_adam.py

 import math
 import torch
 import importlib
+import amp_C
 from apex.multi_tensor_apply import multi_tensor_applier
 
 class DistributedFusedAdam(torch.optim.Optimizer):
@@ -46,9 +47,9 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                  amp_scale_adjustment=1.0, overlap_reductions=True, full_pipeline=True,
                  compute_L2_grad_norm=False, distributed_weight_update=0,
                  dwu_group_size=0, dwu_num_blocks=4, dwu_num_rs_pg=1, dwu_num_ar_pg=4,
-                 dwu_num_ag_pg=0, dwu_num_blk_st=1, revert_method=1, flat_mt=False,
-                 dwu_num_chunks=4, predivide=True, internal_pipeline=False,
-                 e5m2_allgather=False):
+                 dwu_num_ag_pg=0, revert_method=1, flat_mt=False,
+                 dwu_num_chunks=4, predivide=True, e5m2_allgather=False,
+                 do_not_flatten_model=False):
         global fused_adam_cuda
         fused_adam_cuda = importlib.import_module("fused_adam_cuda")
 
@@ -67,6 +68,8 @@ class DistributedFusedAdam(torch.optim.Optimizer):
 
         self._overflow_buf = torch.cuda.IntTensor([0])
 
+        assert (not flat_mt), "flat_mt option is not safe in this version"
+
         # Way to revert a step
         # 3 -> undo kernel + double buffer (debug, print norm of difference)
         # 2 -> double buffer fp32 parameters
@@ -81,8 +84,8 @@ class DistributedFusedAdam(torch.optim.Optimizer):
         self._num_blocks = dwu_num_blocks
         self._num_chunks = dwu_num_chunks
         self._predivide = predivide
-        self._internal_pipeline = internal_pipeline
         self._e5m2_allgather = e5m2_allgather
+        self._do_not_flatten_model = do_not_flatten_model
         self._full_pipeline = full_pipeline
         self._compute_L2_grad_norm = compute_L2_grad_norm
         self._L2_grad_norm = torch.zeros([]).cuda() if self._compute_L2_grad_norm else None
@@ -118,11 +121,12 @@ class DistributedFusedAdam(torch.optim.Optimizer):
 
         self._net_total_param_size = p_offset
         self._total_param_size = p_offset
-        dwu_min_page_size = 256 * self._num_blocks * self._group_size
+        dwu_min_page_size = 256 * self._num_blocks * self._num_chunks * self._group_size
         self._total_param_size = ((self._total_param_size + dwu_min_page_size - 1) // dwu_min_page_size) * dwu_min_page_size
         self._block_size = self._total_param_size // self._num_blocks
-        self._shard_size = self._block_size // self._group_size
-        print("self._net_total_param_size=%d, self._total_param_size=%d, dwu_min_page_size=%d, self._block_size=%d, self._shard_size=%d" % (self._net_total_param_size, self._total_param_size,dwu_min_page_size,self._block_size,self._shard_size))
+        self._chunk_size = self._block_size // self._num_chunks
+        self._shard_size = self._chunk_size // self._group_size
+        print("self._net_total_param_size=%d, self._total_param_size=%d, dwu_min_page_size=%d, self._block_size=%d, self._chunk_size=%d, self._shard_size=%d" % (self._net_total_param_size, self._total_param_size,dwu_min_page_size,self._block_size,self._chunk_size,self._shard_size))
 
         self._low_param_i = [0]*self._num_blocks
         for block_id in range(self._num_blocks-1,-1,-1):
@@ -143,7 +147,6 @@ class DistributedFusedAdam(torch.optim.Optimizer):
         self._num_rs_pg = dwu_num_rs_pg
         self._num_ar_pg = dwu_num_ar_pg
         self._num_ag_pg = dwu_num_ag_pg
-        self._num_blk_st = dwu_num_blk_st
         if self._num_groups > 1:
             self._ar_pg = []
             for dev_i in range(self._group_size):
@@ -152,6 +155,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                     grp = torch.distributed.new_group(ranks=ranks)
                     if torch.distributed.get_rank() in ranks:
                         self._ar_pg.append(grp)
+            self._ar_st = [torch.cuda.Stream()]*self._num_ar_pg
         rs_ranks = []
         for group_i in range(self._num_groups):
             rs_ranks.append([group_i*self._group_size+j for j in range(self._group_size)])
@@ -162,8 +166,13 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                 grp = torch.distributed.new_group(ranks=ranks)
                 if torch.distributed.get_rank() in ranks:
                     self._rs_pg.append(grp)
+            if self._compute_L2_grad_norm and torch.distributed.get_rank() in ranks:
+                self._l2_grad_norm_pg = torch.distributed.new_group(ranks=ranks)
+        self._rs_st = [torch.cuda.Stream()]*self._num_rs_pg
         if self._num_ag_pg == 0:
             self._ag_pg = self._rs_pg
+            self._ag_st = self._rs_st
+            self._num_ag_pg = self._num_rs_pg
         else:
             self._ag_pg = []
             for group_i in range(self._num_groups):
@@ -172,16 +181,15 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                     grp = torch.distributed.new_group(ranks=ranks)
                     if torch.distributed.get_rank() in ranks:
                         self._ag_pg.append(grp)
-        self._blk_st = []
-        for i in range(self._num_blk_st):
-            self._blk_st.append(torch.cuda.Stream())
+            self._ag_st = [torch.cuda.Stream()]*self._num_ag_pg
+        self._l2_grad_norm_st = torch.cuda.Stream() if self._compute_L2_grad_norm else None
+        self._completion_st = torch.cuda.Stream()
+
+        self._reductions_works = [None]*self._num_blocks
+        self._allgather_works = [None]*self._num_blocks
 
         import inspect
-        if 'no_copy' in inspect.getfullargspec(torch.distributed.reduce_scatter).args:
-            self._pg_supports_no_copy = True
-        else:
-            self._pg_supports_no_copy = False
-            print("WARNING! torch.distributed.reduce_scatter does not support no_copy op.")
+        assert ('no_copy' in inspect.getfullargspec(torch.distributed.reduce_scatter).args), "This version of c10d does not support no_copy option"
 
 
     def set_last_step(self, last_step):
@@ -207,71 +215,65 @@ class DistributedFusedAdam(torch.optim.Optimizer):
 
         return flush_block
 
-    def _pipeline_block_reductions(self, block_id, flat_grads):
+    def _pipeline_block_reductions(self, block_id):
+        self._flatten_grad_mt(1.0/self._world_size if self._predivide else 1.0)
+
         start = block_id * self._block_size
         end = start + self._block_size
-        grad_block = flat_grads[start:end]
-        grad_shards = [grad_block[i*self._shard_size:(i+1)*self._shard_size] for i in range(self._group_size)]
-
-        if self._internal_pipeline:
-            works = []
-            chunk_size = self._shard_size // self._num_chunks
-            for i in range(self._num_chunks):
-                chunks = [grad_shards[j][i*chunk_size:(i+1)*chunk_size] for j in range(self._group_size)]
-                if self._pg_supports_no_copy:
-                    work = torch.distributed.reduce_scatter(chunks[self._rank_in_group],chunks,group=self._rs_pg[i%len(self._rs_pg)],async_op=True,no_copy=True)
-                else:
-                    work = torch.distributed.reduce_scatter(chunks[self._rank_in_group],chunks,group=self._rs_pg[i%len(self._rs_pg)],async_op=True)
-                if self._num_groups > 1:
-                    work.wait()
-                    work = torch.distributed.all_reduce(chunks[self._rank_in_group],group=self._ar_pg[i%len(self._ar_pg)],async_op=True)
-                works.append(work)
-        else:
-            if self._pg_supports_no_copy:
-                work = torch.distributed.reduce_scatter(grad_shards[self._rank_in_group],grad_shards,group=self._rs_pg[block_id%len(self._rs_pg)],async_op=True,no_copy=True)
-            else:
-                work = torch.distributed.reduce_scatter(grad_shards[self._rank_in_group],grad_shards,group=self._rs_pg[block_id%len(self._rs_pg)],async_op=True)
-            works = [work]
+        grad_block = self._flat_grads[start:end]
+
+        works = [None]*self._num_chunks
+        for chunk in range(self._num_chunks):
+            grad_chunk = grad_block[chunk*self._chunk_size:(chunk+1)*self._chunk_size]
+            grad_shards = [grad_chunk[i*self._shard_size:(i+1)*self._shard_size] for i in range(self._group_size)]
+            rs_stream = self._rs_st[chunk%self._num_rs_pg]
+            rs_stream.wait_stream(torch.cuda.current_stream())
+            with torch.cuda.stream(rs_stream):
+                work = torch.distributed.reduce_scatter(grad_shards[self._rank_in_group],grad_shards,group=self._rs_pg[chunk%self._num_rs_pg],async_op=True,no_copy=True)
             if self._num_groups > 1:
-                work.wait()
-                works = []
-                chunk_size = self._shard_size // self._num_chunks
-                for i in range(self._num_chunks):
-                    chunks = [grad_shards[j][i*chunk_size:(i+1)*chunk_size] for j in range(self._group_size)]
-                    work = torch.distributed.all_reduce(chunks[self._rank_in_group],group=self._ar_pg[i%len(self._ar_pg)],async_op=True)
-                    works.append(work)
-
-        if self._compute_L2_grad_norm:
-            with torch.cuda.stream(self._blk_st[0]):
-                for work in works:
+                ar_stream = self._ar_st[chunk%self._num_ar_pg]
+                with torch.cuda.stream(ar_stream):
                     work.wait()
-                if block_id+1 == self._num_blocks:
-                    self._L2_grad_norm = grad_shards[self._rank_in_group].norm(dtype=torch.float32,p=2)**2
-                elif block_id != 0:
-                    self._L2_grad_norm += grad_shards[self._rank_in_group].norm(dtype=torch.float32,p=2)**2
-                else:
-                    self._L2_grad_norm += grad_shards[self._rank_in_group].norm(dtype=torch.float32,p=2)**2
-                    torch.distributed.all_reduce(self._L2_grad_norm,group=self._rs_pg[0])
-                    self._L2_grad_norm.sqrt_()
-
-        for work in works:
-            work.wait()
+                    work = torch.distributed.all_reduce(grad_shards[self._rank_in_group],group=self._ar_pg[chunk%self._num_ar_pg],async_op=True)
+            works[chunk] = work
 
-    # NB!
-    # self._global_scale is used by this method.
+        if self._compute_L2_grad_norm:
+            for chunk in range(self._num_chunks):
+                grad_chunk = grad_block[chunk*self._chunk_size:(chunk+1)*self._chunk_size]
+                grad_shards = [grad_chunk[i*self._shard_size:(i+1)*self._shard_size] for i in range(self._group_size)]
+                with torch.cuda.stream(self._l2_grad_norm_st):
+                    works[chunk].wait()
+                    l2_grad_sq = grad_shards[self._rank_in_group].norm(dtype=torch.float32,p=2)**2
+                    if block_id+1 == self._num_blocks and chunk == 0:
+                        self._L2_grad_norm = l2_grad_sq
+                    else:
+                        self._L2_grad_norm += l2_grad_sq
+                    if block_id == 0 and chunk+1 == self._num_chunks:
+                        torch.distributed.all_reduce(self._L2_grad_norm,group=self._l2_grad_norm_pg)
+                        self._L2_grad_norm.sqrt_()
+
+        self._reductions_works[block_id] = works
+
+    def _pipeline_block_step(self, block_id):
+        if self._new_params is None:
+            self._new_params = torch.zeros_like(self._flat_grads,dtype=uint8 if self._e5m2_allgather else self._flat_grads.dtype)
 
-    def _pipeline_block_step(self, block_id, flat_grads, new_params):
         start = block_id * self._block_size
-        new_params_shards = [new_params[start+shard_i*self._shard_size:start+(shard_i+1)*self._shard_size] for shard_i in range(self._group_size)]
-        self._partial_step_single_shard(block_id)
-        if self._pg_supports_no_copy:
-            torch.distributed.all_gather(new_params_shards,new_params_shards[self._rank_in_group],group=self._ag_pg[block_id%len(self._ag_pg)],no_copy=True)
-        else:
-            torch.distributed.all_gather(new_params_shards,new_params_shards[self._rank_in_group],group=self._ag_pg[block_id%len(self._ag_pg)])
+        end = start + self._block_size
+        new_params_block = self._new_params[start:end]
 
-    def _pipeline_block(self, block_id, flat_grads, new_params):
-        self._pipeline_block_reductions(block_id, flat_grads)
-        self._pipeline_block_step(block_id, flat_grads, new_params)
+        works = [None]*self._num_chunks
+        for chunk in range(self._num_chunks):
+            new_params_chunk = new_params_block[chunk*self._chunk_size:(chunk+1)*self._chunk_size]
+            new_params_shards = [new_params_chunk[i*self._shard_size:(i+1)*self._shard_size] for i in range(self._group_size)]
+            ag_stream = self._ag_st[chunk%self._num_ag_pg]
+            with torch.cuda.stream(ag_stream):
+                self._reductions_works[block_id][chunk].wait()
+                self._partial_step_single_shard(block_id,chunk)
+                work = torch.distributed.all_gather(new_params_shards,new_params_shards[self._rank_in_group],group=self._ag_pg[chunk%self._num_ag_pg],async_op=True,no_copy=True)
+                works[chunk] = work
+
+        self._allgather_works[block_id] = works
 
     def _flatten_grad_mt(self, scale):
         if self._flat_mt:
@@ -281,10 +283,8 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                 if grad is not None:
                     grads.append(grad)
                     flat_grads.append( self._flat_grads[grads_info["param_offset"]:grads_info["param_offset"]+grads_info["param_grads_size"]] )
-                    self._grads[p_i] = None
+            self._grads = [None]*len(self._grads_info)
             if len(grads) > 0:
-                import amp_C
-                from apex.multi_tensor_apply import multi_tensor_applier
                 self._overflow_buf.zero_()
                 multi_tensor_applier(
                         amp_C.multi_tensor_scale,
@@ -295,7 +295,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
     def _do_overlapped_reduction(self, param_i, param_grads_size, param_offset, grad):
         # handle overlapped reductions
         if self._flat_mt:
-            self._grads[param_i] = grad
+            self._grads[param_i] = grad.view(-1)
         else:
             torch.div(grad.view(-1), self._world_size if self._predivide else 1.0, out=self._flat_grads[param_offset:param_offset+param_grads_size])
         self._grads_generated[param_i]=True
@@ -304,19 +304,9 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                 flush_block = self._get_flush_block()
                 while flush_block:
                     block_id = flush_block[0] // self._block_size
-                    self._flatten_grad_mt(1.0/self._world_size if self._predivide else 1.0)
-                    self._blk_st[block_id%len(self._blk_st)].wait_stream(torch.cuda.current_stream())
-                    with torch.cuda.stream(self._blk_st[block_id%len(self._blk_st)]):
-                        if self._full_pipeline:
-                            if self._new_params is None:
-                                if self._e5m2_allgather:
-                                    self._new_params = torch.zeros_like(self._flat_grads,dtype=torch.uint8)
-                                else:
-                                    self._new_params = torch.zeros_like(self._flat_grads)
-                            self._pipeline_block(block_id, self._flat_grads, self._new_params)
-                        else:
-                            self._pipeline_block_reductions(block_id, self._flat_grads)
-
+                    self._pipeline_block_reductions(block_id)
+                    if self._full_pipeline:
+                        self._pipeline_block_step(block_id)
                     flush_block = self._get_flush_block()
 
     def set_global_scale(self, global_scale):
@@ -360,8 +350,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
     @property
     def L2_grad_norm(self):
         if self._compute_L2_grad_norm:
-            for i, blk_st in enumerate(self._blk_st):
-                torch.cuda.current_stream().wait_stream(blk_st)
+            torch.cuda.current_stream().wait_stream(self._l2_grad_norm_st)
             return self._L2_grad_norm
         else:
             return None
@@ -376,7 +365,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
     # This means we have to play around with indexes, which requires knowledge of block and shard number.
     # Implement a method that performs a partial update of a single shard within a single block.
 
-    def _partial_step_single_shard(self, block_id, undo=False):
+    def _partial_step_single_shard(self, block_id, chunk_id, undo=False):
         """Perform step function for a single shard.
 
         Arguments:
@@ -385,7 +374,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
 
         """
         shard_id = self._rank_in_group
-        shard_start = block_id * self._block_size + shard_id * self._shard_size
+        shard_start = (block_id * self._num_chunks + chunk_id) * self._chunk_size + shard_id * self._shard_size
         shard_end = shard_start + self._shard_size
 
         if self._fp32_p is None:
@@ -393,13 +382,13 @@ class DistributedFusedAdam(torch.optim.Optimizer):
             # Allocate fp32 buffers on demand. Note that we don't make these part of the state
             # since each rank only has partial buffers.
             # To-Do: 
-            self._fp32_p = torch.zeros([self._num_blocks*self._shard_size]).float().cuda()
-            self._fp32_m = torch.zeros([self._num_blocks*self._shard_size]).float().cuda()
-            self._fp32_v = torch.zeros([self._num_blocks*self._shard_size]).float().cuda()
+            self._fp32_p = torch.zeros([self._num_blocks*self._num_chunks*self._shard_size]).float().cuda()
+            self._fp32_m = torch.zeros([self._num_blocks*self._num_chunks*self._shard_size]).float().cuda()
+            self._fp32_v = torch.zeros([self._num_blocks*self._num_chunks*self._shard_size]).float().cuda()
             if self._revert_method > 1:
-                self._fp32_backup_p = torch.zeros([self._num_blocks*self._shard_size]).float().cuda()
-                self._fp32_backup_m = torch.zeros([self._num_blocks*self._shard_size]).float().cuda()
-                self._fp32_backup_v = torch.zeros([self._num_blocks*self._shard_size]).float().cuda()
+                self._fp32_backup_p = torch.zeros([self._num_blocks*self._num_chunks*self._shard_size]).float().cuda()
+                self._fp32_backup_m = torch.zeros([self._num_blocks*self._num_chunks*self._shard_size]).float().cuda()
+                self._fp32_backup_v = torch.zeros([self._num_blocks*self._num_chunks*self._shard_size]).float().cuda()
             self._copy_to_fp32 = True
 
         step = None
@@ -445,7 +434,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                     if self._copy_to_fp32:
                         param_offset = clipped_start - shard_start
                         param_size = clipped_end - clipped_start
-                        buffer_start = block_id * self._shard_size + param_offset
+                        buffer_start = (block_id * self._num_chunks + chunk_id) * self._shard_size + param_offset
                         buffer_end = buffer_start + param_size
                         param_start = (clipped_start - start)
                         param_end = param_start + param_size
@@ -457,7 +446,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                 group_offset = group_start - shard_start
                 group_shard_start = shard_start + group_offset
                 group_shard_end = group_shard_start + group_size
-                group_buffer_start = block_id * self._shard_size + group_offset
+                group_buffer_start = (block_id * self._num_chunks + chunk_id) * self._shard_size + group_offset
                 group_buffer_end = group_buffer_start + group_size
 
                 beta1, beta2 = group['betas']
@@ -520,12 +509,11 @@ class DistributedFusedAdam(torch.optim.Optimizer):
 
         if self._last_step or not self._overlap_reductions:
             # nothing done so far, run full pipeline after reductions
-            for inv_block_id in range(self._num_blocks):
-                block_id = self._num_blocks - inv_block_id - 1
-                self._flatten_grad_mt(1.0/self._world_size if self._predivide else 1.0)
-                self._blk_st[block_id%len(self._blk_st)].wait_stream(torch.cuda.current_stream())
-                with torch.cuda.stream(self._blk_st[block_id%len(self._blk_st)]):
-                    self._pipeline_block_reductions(block_id, self._flat_grads)
+            for block_id in range(self._num_blocks-1,-1,-1):
+                self._pipeline_block_reductions(block_id)
+
+        if self._compute_L2_grad_norm:
+            torch.cuda.current_stream().wait_stream(self._l2_grad_norm_st)
 
         self._copy_to_fp32 = False
         self._decomp_stats = None
@@ -536,7 +524,8 @@ class DistributedFusedAdam(torch.optim.Optimizer):
         """Revert effect of previously calling partial_step.
         """
         for block_id in range(self._num_blocks):
-            self._partial_step_single_shard(block_id, undo=True)
+            for chunk in range(self._num_chunks):
+                self._partial_step_single_shard(block_id, chunk, undo=True)
 
     def step(self, closure=None, skip_overflow_check=False):
         loss = None
@@ -544,19 +533,13 @@ class DistributedFusedAdam(torch.optim.Optimizer):
             loss = closure()
 
         if self._last_step or not self._overlap_reductions or not self._full_pipeline:
-            if self._new_params is None:
-                if self._e5m2_allgather:
-                    self._new_params = torch.zeros_like(self._flat_grads,dtype=torch.uint8)
-                else:
-                    self._new_params = torch.zeros_like(self._flat_grads)
-            for inv_block_id in range(self._num_blocks):
-                block_id = self._num_blocks - inv_block_id - 1
-                with torch.cuda.stream(self._blk_st[block_id%len(self._blk_st)]):
-                    self._pipeline_block_step(block_id, self._flat_grads, self._new_params)
+            for block_id in range(self._num_blocks-1,-1,-1):
+                self._pipeline_block_step(block_id)
 
-        with torch.cuda.stream(self._blk_st[0]):
-            for i, blk_st in enumerate(self._blk_st):
-                torch.cuda.current_stream().wait_stream(blk_st)
+        with torch.cuda.stream(self._completion_st):
+            for block_id in range(self._num_blocks-1,-1,-1):
+                for chunk in range(self._num_chunks):
+                    self._allgather_works[block_id][chunk].wait()
 
             # Check for overflow
             # Store state for loss scaler calculation
@@ -570,7 +553,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                 self.revert_step()
             else:
                 # Copy self._new_params to model params
-                if self._e5m2_allgather:
+                if self._e5m2_allgather or self._do_not_flatten_model:
                     p_in = []
                     p_out = []
                 with torch.no_grad():
@@ -585,7 +568,7 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                             state['step'] += 1
                             nels = p.numel()
                             offset = self._grads_info[param_i]['param_offset']
-                            if self._e5m2_allgather:
+                            if self._e5m2_allgather or self._do_not_flatten_model:
                                 p_in.append(self._new_params[offset:offset+nels].view_as(p))
                                 p_out.append(p)
                             else:
@@ -596,9 +579,20 @@ class DistributedFusedAdam(torch.optim.Optimizer):
                                 fused_adam_cuda.unpack_e5m2_mt,
                                 self._overflow_buf,
                                 [p_in, p_out]);
-            self._new_params = None
+                    elif self._do_not_flatten_model:
+                        multi_tensor_applier(
+                                amp_C.multi_tensor_scale,
+                                self._overflow_buf,
+                                [p_in, p_out],
+                                1.0);
+            if not self._e5m2_allgather and not self._do_not_flatten_model:
+                self._new_params = None
+
+        torch.cuda.current_stream().wait_stream(self._completion_st)
+
+        self._reductions_works = [None]*self._num_blocks
+        self._allgather_works = [None]*self._num_blocks
 
-        torch.cuda.current_stream().wait_stream(self._blk_st[0])
         return loss