Skip to content

fix(training): align DSv4 packed FLOPs with MCore #5542

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

Merged
yaoyu-33 merged 1 commit into from
Aug 14, 2026
+105 −7
Merged

fix(training): align DSv4 packed FLOPs with MCore #5542

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
21 changes: 14 additions & 7 deletions src/megatron/bridge/training/utils/flop_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -668,8 +668,10 @@ def num_floating_point_operations(
# the result matches the legacy constant-length estimate.
if seqlen_squared_sum is not None and seqlen_sum > 0:
core_attn_seq_factor = seqlen_squared_sum / seqlen_sum
effective_seqlen_squared_sum = seqlen_squared_sum
else:
core_attn_seq_factor = effective_seq_length
effective_seqlen_squared_sum = seqlen_sum * effective_seq_length

# If the model provider has a custom TFLOPS calculation method, use it (non-LoRA only).
if not is_lora and hasattr(cfg.model, "_get_num_floating_point_operations"):
Expand Down Expand Up @@ -1049,6 +1051,7 @@ def transformer_flops():
ffn_expansion_factor = 3 if cfg.model.gated_linear_unit is True else 2

experimental_attention_variant = getattr(cfg.model, "experimental_attention_variant", None)
dsv4_hybrid_core_attn_term = 0

if cfg.model.multi_latent_attention:
"""
Expand Down Expand Up @@ -1117,10 +1120,9 @@ def transformer_flops():
window = getattr(cfg.model, "csa_window_size", 128)

sparse_attn_r0 = n_layers_r0 * cfg.model.num_attention_heads * window * v_head_dim * 2
avg_comp_128 = (core_attn_seq_factor // 128) / 2
sparse_attn_r128 = (
n_layers_r128 * cfg.model.num_attention_heads * (window + avg_comp_128) * v_head_dim * 2
)
# Window work is token-linear; compressed-KV attention scales with sum_i(sequence_length_i^2).
sparse_attn_r128 = n_layers_r128 * cfg.model.num_attention_heads * window * v_head_dim * 2
sparse_attn_r128_core = n_layers_r128 * cfg.model.num_attention_heads * v_head_dim / 128

main_compressor_term = (
n_layers_r4 * cfg.model.hidden_size * (2 * v_head_dim) * 2
Expand All @@ -1138,23 +1140,27 @@ def transformer_flops():
if idx_topk is None:
raise ValueError("dsa_indexer_topk must be set for dsv4_hybrid ratio==4 layers")

effective_topk_4 = min(idx_topk, core_attn_seq_factor // 4)
avg_comp_4 = effective_topk_4 * (1 - effective_topk_4 * 4 / (2 * core_attn_seq_factor))
# Match MCore's nominal ratio-4 selection estimate, which uses the configured sequence length.
effective_topk_4 = min(idx_topk, cfg.model.seq_length // 4)
avg_comp_4 = effective_topk_4 * (1 - effective_topk_4 * 4 / (2 * cfg.model.seq_length))
sparse_attn_r4 = (
n_layers_r4 * cfg.model.num_attention_heads * (window + avg_comp_4) * v_head_dim * 2
)
indexer_term = (
n_layers_r4 * cfg.model.hidden_size * (2 * idx_head_dim) * 2
+ n_layers_r4 * q_lora_rank * idx_n_heads * idx_head_dim
+ n_layers_r4 * cfg.model.hidden_size * idx_n_heads
+ n_layers_r4 * idx_n_heads * idx_head_dim * (core_attn_seq_factor // 4)
)
# Dense indexer scoring is quadratic and therefore uses the runtime squared-length sum below.
indexer_scoring_core = n_layers_r4 * idx_n_heads * idx_head_dim / 4
else:
sparse_attn_r4 = 0
indexer_term = 0
indexer_scoring_core = 0

sparse_attn_term = sparse_attn_r0 + sparse_attn_r4 + sparse_attn_r128
self_attn_term += 3 * 2 * (sparse_attn_term + main_compressor_term + indexer_term)
dsv4_hybrid_core_attn_term = 3 * 2 * (sparse_attn_r128_core + indexer_scoring_core)
elif experimental_attention_variant == "dsa":
# DSA replaces dense MLA core attention with top-k attention while retaining a
# dense lightning indexer. The attention/indexer geometry follows equations 1-2
Expand Down Expand Up @@ -1556,6 +1562,7 @@ def count_indexer_layers(layer_count: int) -> int:
# Logit.
+ 3 * 2 * cfg.model.hidden_size * padded_vocab_size * (mtp_num_layers + 1)
)
total_floating_point_operations += effective_seqlen_squared_sum * dsv4_hybrid_core_attn_term
return total_floating_point_operations + _compute_vit_flops()

def _compute_vit_flops():
Expand Down
91 changes: 91 additions & 0 deletions tests/unit_tests/training/utils/test_flop_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1057,6 +1057,97 @@ def test_dsv4_hybrid_exact_flops(self):

assert actual_flops == expected_flops

def test_dsv4_hybrid_packed_flops_match_mcore_split(self):
"""Packed DSv4 FLOPs split token-linear work from quadratic sparse work."""
batch_size = 2
seq_len = 256
hidden_size = 512
num_layers = 4
num_heads = 8
v_head_dim = 64
q_lora_rank = 128
o_lora_rank = 64
o_groups = 2
window = 64
idx_n_heads = 4
idx_head_dim = 32
idx_topk = 16
ffn_hidden_size = 2048
vocab_size = 1024
compress_ratios = [0, 4, 128, 128]
packed_lengths = [64, 64, 128, 256]
seqlen_sum = sum(packed_lengths)
seqlen_squared_sum = sum(length**2 for length in packed_lengths)

model_cfg = MockModelConfig(
num_layers=num_layers,
hidden_size=hidden_size,
seq_length=seq_len,
ffn_hidden_size=ffn_hidden_size,
num_attention_heads=num_heads,
vocab_size=vocab_size,
multi_latent_attention=True,
experimental_attention_variant="dsv4_hybrid",
q_lora_rank=q_lora_rank,
qk_head_dim=32,
qk_pos_emb_head_dim=32,
v_head_dim=v_head_dim,
o_lora_rank=o_lora_rank,
o_groups=o_groups,
csa_compress_ratios=compress_ratios,
csa_window_size=window,
dsa_indexer_n_heads=idx_n_heads,
dsa_indexer_head_dim=idx_head_dim,
dsa_indexer_topk=idx_topk,
gated_linear_unit=False,
)
cfg = MockConfigContainer(model=model_cfg)

q_term = q_lora_rank * (hidden_size + num_heads * v_head_dim + 1)
kv_term = hidden_size * v_head_dim + v_head_dim
o_term = num_heads * v_head_dim * o_lora_rank + o_groups * o_lora_rank * hidden_size
projection_term = 3 * 2 * num_layers * (q_term + kv_term + o_term)

n_layers_r0 = compress_ratios.count(0)
n_layers_r4 = compress_ratios.count(4)
n_layers_r128 = compress_ratios.count(128)
sparse_attn_r0 = n_layers_r0 * num_heads * window * v_head_dim * 2
sparse_attn_r128_window = n_layers_r128 * num_heads * window * v_head_dim * 2
effective_topk = min(idx_topk, seq_len // 4)
avg_comp_4 = effective_topk * (1 - effective_topk * 4 / (2 * seq_len))
sparse_attn_r4 = n_layers_r4 * num_heads * (window + avg_comp_4) * v_head_dim * 2
main_compressor_term = (
n_layers_r4 * hidden_size * (2 * v_head_dim) * 2 + n_layers_r128 * hidden_size * v_head_dim * 2
)
indexer_token_term = (
n_layers_r4 * hidden_size * (2 * idx_head_dim) * 2
+ n_layers_r4 * q_lora_rank * idx_n_heads * idx_head_dim
+ n_layers_r4 * hidden_size * idx_n_heads
)
self_attn_token_term = projection_term + 3 * 2 * (
sparse_attn_r0 + sparse_attn_r4 + sparse_attn_r128_window + main_compressor_term + indexer_token_term
)
self_attn_core_term = (
3 * 2 * (n_layers_r128 * num_heads * v_head_dim / 128 + n_layers_r4 * idx_n_heads * idx_head_dim / 4)
)

mlp_term = 3 * 2 * hidden_size * (ffn_hidden_size * 2 * num_layers)
logit_term = 3 * 2 * hidden_size * vocab_size
expected_flops = (
seqlen_sum * (mlp_term + self_attn_token_term + logit_term) + seqlen_squared_sum * self_attn_core_term
)

actual_flops = num_floating_point_operations(
cfg,
batch_size=batch_size,
seqlen_sum=seqlen_sum,
seqlen_squared_sum=seqlen_squared_sum,
)
bshd_flops = num_floating_point_operations(cfg, batch_size=batch_size)

assert actual_flops == expected_flops
assert actual_flops < bshd_flops

def test_dsv4_hybrid_validates_compress_ratio_length(self):
"""CSA compress-ratio count must match decoder plus MTP layer count."""
model_cfg = MockModelConfig(
Expand Down
Loading