Skip to content

refactor(loader): gate batched-prefill MLA exclusion on capability, not model_type #261

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
SeedSource merged 1 commit into from
Jul 9, 2026
Merged

refactor(loader): gate batched-prefill MLA exclusion on capability, not model_type #261

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
16 changes: 11 additions & 5 deletions crates/spark-model/src/model/trait_impl/prefill_b/batch_kernel/eligible.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,7 +65,10 @@ impl TransformerModel {
.map(|s| (s.chunk_len, s.chunk_start, s.is_last_chunk)),
streams.len(),
self.buffers.max_batch_tokens(),
&self.config.model_type,
// is_mla == AttentionType::Mla; read the underlying fact directly to
// avoid rebuilding the full ModelCapabilities struct on this
// per-prefill-batch hot path.
self.config.kv_lora_rank > 0,
self.config.head_dim,
self.buffers.scratch_bytes(),
self.config.num_experts_per_tok,
Expand Down Expand Up @@ -139,7 +142,7 @@ pub(in crate::model) fn check_kernel_batched_eligible<I>(
streams: I,
n: usize,
arena_cap: usize,
model_type: &str,
is_mla: bool,
head_dim: usize,
scratch_cap: usize,
top_k: usize,
Expand All @@ -154,9 +157,12 @@ where
return false;
}
// No MLA layers in stack (batched attention doesn't support MLA).
// Conservatively check via model_type — mistral is the only MLA
// model in Atlas today.
if model_type == "mistral" {
// Keyed on the MLA capability (AttentionType::Mla, i.e. kv_lora_rank>0),
// which is architecture-generic: it covers mistral AND deepseek_v4 (the
// old `model_type=="mistral"` string silently missed the latter, though
// deepseek_v4 was already rejected one check later by head_dim>256, so the
// outcome is unchanged on every model shipped today).
if is_mla {
return false;
}
// No HDIM=512 layers (Gemma-4 long-attention).
Expand Down
28 changes: 14 additions & 14 deletions crates/spark-model/src/model/trait_impl/prefill_b/batch_kernel_tests.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@ fn rejects_under_two_streams() {
std::iter::empty(),
0,
8192,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -36,7 +36,7 @@ fn rejects_under_two_streams() {
vec![s(4096, 0, false)],
1,
8192,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -52,7 +52,7 @@ fn rejects_chunk_zero() {
vec![s(4096, 0, false), s(4096, 0, false)],
2,
8192,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -68,7 +68,7 @@ fn accepts_chunk_zero_when_explicitly_allowed() {
vec![s(4096, 0, false), s(4096, 0, false)],
2,
8192,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -84,7 +84,7 @@ fn accepts_uniform_paged_n_2() {
vec![s(4096, 4096, false), s(4096, 4096, false)],
2,
8192,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -100,7 +100,7 @@ fn rejects_mismatched_chunk_len() {
vec![s(4096, 4096, false), s(2048, 4096, false)],
2,
16384,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -118,7 +118,7 @@ fn rejects_mismatched_chunk_start() {
vec![s(4096, 12288, false), s(4096, 4096, false)],
2,
16384,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -134,7 +134,7 @@ fn rejects_mismatched_is_last() {
vec![s(4096, 4096, false), s(4096, 4096, true)],
2,
8192,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -151,7 +151,7 @@ fn rejects_arena_overflow() {
vec![s(4096, 4096, false), s(4096, 4096, false)],
2,
4100,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -167,7 +167,7 @@ fn rejects_mla_model() {
vec![s(4096, 4096, false), s(4096, 4096, false)],
2,
8192,
"mistral",
true,
128,
BIG_SCRATCH,
TOP_K,
Expand All @@ -184,7 +184,7 @@ fn rejects_large_head_dim() {
vec![s(4096, 4096, false), s(4096, 4096, false)],
2,
8192,
"gemma4",
false,
512,
BIG_SCRATCH,
TOP_K,
Expand All @@ -200,7 +200,7 @@ fn accepts_n_4_uniform() {
vec![s(2048, 2048, false); 4],
4,
8192,
"qwen3_next",
false,
256,
BIG_SCRATCH,
TOP_K,
Expand All @@ -227,7 +227,7 @@ fn rejects_scratch_footprint_overflow() {
streams.iter().copied(),
4,
arena,
"qwen3_next",
false,
256,
too_small,
8,
Expand All @@ -242,7 +242,7 @@ fn rejects_scratch_footprint_overflow() {
streams.iter().copied(),
4,
arena,
"qwen3_next",
false,
256,
enlarged,
8,
Expand Down
Loading