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 new optimizer state row_counter for Adam [Backend] #3342

Closed
wants to merge 1 commit into from
Closed

Add new optimizer state row_counter for Adam [Backend] #3342

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
353 changes: 284 additions & 69 deletions fbgemm_gpu/codegen/genscript/optimizer_args.py
View file
Open in desktop

Large diffs are not rendered by default.

37 changes: 31 additions & 6 deletions fbgemm_gpu/codegen/genscript/optimizers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1001,6 +1001,29 @@ def partial_rowwise_lamb() -> Dict[str, Any]:


def adam() -> Dict[str, Any]:
split_precomputation = """
at::acc_type<cache_t, true>* __restrict__ row_counter;
at::acc_type<cache_t, true> _row_counter = iter;
if (use_rowwise_bias_correction) {
const auto row_counter_placement = static_cast<PlacementType>(row_counter_placements[t]);
const int64_t row_counter_offset = row_counter_offsets[t];
if (row_counter_placement == PlacementType::DEVICE) {
row_counter = &row_counter_dev[row_counter_offset];
} else {
row_counter = &row_counter_uvm[row_counter_offset];
}

// need to compute bias correction for each row
if (threadIdx.x == 0) {
_row_counter = row_counter[idx] + 1;
row_counter[idx] = _row_counter;
}

// broadcast bias correction to all threads
_row_counter = SHFL_SYNC(_row_counter, 0);
}
"""

split_weight_update = """
Vec4T<cache_t> m_t(&momentum1[idx * D + d]);
m_t.acc.x *= beta1;
Expand All @@ -1023,10 +1046,10 @@ def adam() -> Dict[str, Any]:
v_t.fma_(grad, 1.0 - beta2);
v_t.store(&momentum2[idx * D + d]);

weight_new.acc.x -= learning_rate * (m_t.acc.x / (1.0 - powf(beta1, iter)) / (sqrtf((v_t.acc.x / (1.0 - powf(beta2, iter)))) + eps) + weight_decay * weight_new.acc.x);
weight_new.acc.y -= learning_rate * (m_t.acc.y / (1.0 - powf(beta1, iter)) / (sqrtf((v_t.acc.y / (1.0 - powf(beta2, iter)))) + eps) + weight_decay * weight_new.acc.y);
weight_new.acc.z -= learning_rate * (m_t.acc.z / (1.0 - powf(beta1, iter)) / (sqrtf((v_t.acc.z / (1.0 - powf(beta2, iter)))) + eps) + weight_decay * weight_new.acc.z);
weight_new.acc.w -= learning_rate * (m_t.acc.w / (1.0 - powf(beta1, iter)) / (sqrtf((v_t.acc.w / (1.0 - powf(beta2, iter)))) + eps) + weight_decay * weight_new.acc.w);
weight_new.acc.x -= learning_rate * (m_t.acc.x / (1.0 - powf(beta1, _row_counter)) / (sqrtf((v_t.acc.x / (1.0 - powf(beta2, _row_counter)))) + eps) + weight_decay * weight_new.acc.x);
weight_new.acc.y -= learning_rate * (m_t.acc.y / (1.0 - powf(beta1, _row_counter)) / (sqrtf((v_t.acc.y / (1.0 - powf(beta2, _row_counter)))) + eps) + weight_decay * weight_new.acc.y);
weight_new.acc.z -= learning_rate * (m_t.acc.z / (1.0 - powf(beta1, _row_counter)) / (sqrtf((v_t.acc.z / (1.0 - powf(beta2, _row_counter)))) + eps) + weight_decay * weight_new.acc.z);
weight_new.acc.w -= learning_rate * (m_t.acc.w / (1.0 - powf(beta1, _row_counter)) / (sqrtf((v_t.acc.w / (1.0 - powf(beta2, _row_counter)))) + eps) + weight_decay * weight_new.acc.w);
"""
split_weight_update_cpu = "" # TODO

Expand All @@ -1043,12 +1066,14 @@ def adam() -> Dict[str, Any]:
OptimItem(ArgType.FLOAT, "beta2"),
OptimItem(ArgType.FLOAT, "weight_decay"),
OptimItem(ArgType.INT, "iter"),
OptimItem(ArgType.BOOL, "use_rowwise_bias_correction"),
OptimItem(ArgType.TENSOR, "row_counter", is_optional=True),
],
{
"v1": "Tensor momentum1, Tensor momentum2, float learning_rate = 0, float eps = 0, float beta1 = 0, float beta2 = 0, float weight_decay = 0, int iter = 0"
"v1": "Tensor momentum1, Tensor momentum2, float learning_rate = 0, float eps = 0, float beta1 = 0, float beta2 = 0, float weight_decay = 0, int iter = 0, bool use_rowwise_bias_correction = False, Tensor? row_counter = None",
},
),
"split_precomputation": "",
"split_precomputation": split_precomputation,
"split_weight_update": split_weight_update,
"split_post_update": "",
"split_weight_update_cpu": split_weight_update_cpu,
Expand Down
1 change: 1 addition & 0 deletions fbgemm_gpu/codegen/genscript/torch_type_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@ class ArgType(IntEnum):
INT = 7
FLOAT = 8
SYM_INT = 9
BOOL = 10


@dataclass
Expand Down
6 changes: 3 additions & 3 deletions fbgemm_gpu/codegen/training/backward/embedding_backward_split_host_cpu_template.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,14 +64,14 @@ class SplitLookupFunction_{{ optimizer }}_Op : public torch::autograd::Function<
bool gradient_clipping,
double max_gradient,
bool stochastic_rounding,
{{ args.split_function_args | join(", ") }},
{{ args.split_function_args_autograd | join(", ") }},
int64_t output_dtype = static_cast<int64_t>(SparseType::FP32)) {
Tensor indice_weights_value = indice_weights.value_or(Tensor());
Tensor feature_requires_grad_value =
feature_requires_grad.value_or(Tensor());
ctx->save_for_backward({
host_weights, weights_placements, weights_offsets, D_offsets, hash_size_cumsum,
indices, offsets, indice_weights_value, feature_requires_grad_value, {{ args.split_saved_tensors | join(", ") }} });
indices, offsets, indice_weights_value, feature_requires_grad_value, {{ args.split_saved_tensors_optional | join(", ") }} });

ctx->saved_data["total_D"] = total_D;
ctx->saved_data["max_D"] = max_D;
Expand Down Expand Up @@ -242,7 +242,7 @@ Tensor split_embedding_codegen_lookup_{{ optimizer }}_function_cpu(
gradient_clipping,
max_gradient,
stochastic_rounding,
{{ args.split_function_arg_names | join(", ") }},
{{ args.split_function_arg_names_autograd | join(", ") }},
output_dtype)[0];
{% else %}
TORCH_CHECK(false, "split_embedding_codegen_lookup_{{ optimizer }}_function_cpu is deprecated. Please see https://github.com/pytorch/FBGEMM/discussions/1727 for more detail.");
Expand Down
6 changes: 3 additions & 3 deletions fbgemm_gpu/codegen/training/backward/embedding_backward_split_host_template.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -363,7 +363,7 @@ enum SSDTensor {
{%- if ssd %}
ssd_tensors.value(),
{%- endif %}
{{ args.split_function_arg_names | join(", ") }}
{{ args.split_function_arg_names_autograd | join(", ") }}
{%- endif %}
)[0];
{%- endmacro %}
Expand Down Expand Up @@ -623,7 +623,7 @@ class {{ autograd_func }} :
{%- if ssd %}
const at::TensorList& ssd_tensors,
{%- endif %}
{{ args.split_function_args | join(", ") }}
{{ args.split_function_args_autograd | join(", ") }}
{%- else %}
{%- if vbe %}
const std::optional<Tensor>& B_offsets,
Expand Down Expand Up @@ -762,7 +762,7 @@ class {{ autograd_func }} :
ssd_tensors[SSDTensor::{{ tensor | upper }}],
{%- endfor %}
{%- endif %}
{{ args.split_saved_tensors | join(", ") }}
{{ args.split_saved_tensors_optional | join(", ") }}
});

{%- if not nobag %}
Expand Down
44 changes: 35 additions & 9 deletions fbgemm_gpu/codegen/training/pt2/embedding_split_host_pt2_autograd_template.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -476,12 +476,35 @@ enum SSDTensor {

/* This macro generates a code blob for unpacking the tensor list
*/
{%- macro unpack_tensor_list(tensor_list) %}
const Tensor {{ tensor_list }}_host = {{ tensor_list }}[0];
const Tensor {{ tensor_list }}_dev = {{ tensor_list }}[1];
const Tensor {{ tensor_list }}_uvm = {{ tensor_list }}[2];
const Tensor {{ tensor_list }}_placements = {{ tensor_list }}[3];
const Tensor {{ tensor_list }}_offsets = {{ tensor_list }}[4];
{%- macro unpack_tensorlist(name) %}
const Tensor {{ name }}_host = {{ name }}[0];
const Tensor {{ name }}_dev = {{ name }}[1];
const Tensor {{ name }}_uvm = {{ name }}[2];
const Tensor {{ name }}_placements = {{ name }}[3];
const Tensor {{ name }}_offsets = {{ name }}[4];
{%- endmacro %}

{%- macro unpack_tensorlist_optional(name) %}
Tensor {{ name }}_host;
Tensor {{ name }}_dev;
Tensor {{ name }}_uvm;
Tensor {{ name }}_placements;
Tensor {{ name }}_offsets;
if ({{ name }}.has_value()) {
at::TensorList _{{ name }} = {{ name }}.value();
{{ name }}_host = _{{ name }}[0];
{{ name }}_dev = _{{ name }}[1];
{{ name }}_uvm = _{{ name }}[2];
{{ name }}_placements = _{{ name }}[3];
{{ name }}_offsets = _{{ name }}[4];
}
else{
{{ name }}_host = at::empty({0}, weights_host.options());
{{ name }}_dev = at::empty({0}, weights_dev.options());
{{ name }}_uvm = at::empty({0}, weights_uvm.options());
{{ name }}_placements = at::empty({0}, weights_placements.options());
{{ name }}_offsets = at::empty({0}, weights_offsets.options());
}
{%- endmacro %}


Expand Down Expand Up @@ -582,9 +605,12 @@ class {{ autograd_func }} :
{{ args_pt2.unified_pt2.split_function_args | join(", ") }}) {

// unpack Tensor lists
{{ unpack_tensor_list("weights") }}
{%- for arg_name in args_pt2.unified_pt2.split_saved_tensor_list %}
{{ unpack_tensor_list(arg_name) }}
{{ unpack_tensorlist("weights") }}
{%- for arg_name in args_pt2.unified_pt2.split_saved_tensorlist %}
{{ unpack_tensorlist(arg_name) }}
{%- endfor %}
{%- for arg_name in args_pt2.unified_pt2.split_saved_tensorlist_optional %}
{{ unpack_tensorlist_optional(arg_name) }}
{%- endfor %}

const auto T = weights_offsets.sym_numel();
Expand Down
6 changes: 3 additions & 3 deletions fbgemm_gpu/codegen/training/python/split_embedding_codegen_lookup_invoker.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,7 @@ def invoke(
{%- if "prev_iter_dev" in args.split_function_arg_names %}
prev_iter: Momentum,
{%- endif %}
{%- if "row_counter_dev" in args.split_function_arg_names %}
{%- if "row_counter_dev" in args.split_function_arg_names and "row_counter" not in args_pt2.unified_pt2.split_saved_tensorlist_optional %}
row_counter: Momentum,
{%- endif %}
{%- if "iter" in args.split_function_arg_names %}
Expand Down Expand Up @@ -209,7 +209,7 @@ def invoke(
prev_iter_placements=prev_iter.placements,
{%- endif %}
# row_counter
{%- if "row_counter_dev" in args.split_function_arg_names %}
{%- if "row_counter_dev" in args.split_function_arg_names and "row_counter" not in args_pt2.unified_pt2.split_saved_tensorlist_optional %}
row_counter_host=row_counter.host,
row_counter_offsets=row_counter.offsets,
row_counter_placements=row_counter.placements,
Expand Down Expand Up @@ -387,7 +387,7 @@ def invoke(
prev_iter_dev=prev_iter_dev,
{%- endif %}
# row_counter
{%- if "row_counter_dev" in args.split_function_arg_names %}
{%- if "row_counter_dev" in args.split_function_arg_names and "row_counter" not in args_pt2.unified_pt2.split_saved_tensorlist_optional %}
row_counter_dev=row_counter.dev,
row_counter_uvm=row_counter.uvm,
row_counter_offsets=row_counter.offsets,
Expand Down
7 changes: 7 additions & 0 deletions fbgemm_gpu/test/tbe/training/forward_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,6 +76,13 @@
"test_faketensor__test_forward_gpu_uvm_cache_int8": [
unittest.skip("Operator not implemented for Meta tensors"),
],
# learning rate tensor needs to be on CPU to avoid D->H sync point since it will be used as float in the kernel
# this fails fake_tensor test as the test expects all tensors to be on the same device
"test_pt2_compliant_tag_fbgemm_split_embedding_codegen_lookup_rowwise_adagrad_function": [
unittest.skip(
"Operator failed on FakeTensor test since learning rate tensor is always on CPU regardless of other tensors"
),
],
}
)

Expand Down
Loading