Skip to content

perf(gfx12): disable slow default DFlash verify graph #708

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

Open
HUSRCF wants to merge 1 commit into
base: master
Choose a base branch
from
Open

perf(gfx12): disable slow default DFlash verify graph #708

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
23 changes: 19 additions & 4 deletions crates/hipfire-arch-qwen35/src/speculative.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -208,10 +208,10 @@ fn dflash_moe_draft_ffn_graph_eligible(
/// `verify_graph_ok` shares that eligibility via `prefill_batch_pbs_eligible`.
/// Graph-off direct and forced-blob direct full-model V2 fixtures pass on
/// gfx1100, but two graph-on V2 campaigns lost the endpoint. This is a
/// default-off graph capability quarantine for exact gfx1100 + V2 only;
/// direct batched WMMA remains on. `HIPFIRE_VERIFY_GRAPH=1` opts back in
/// diagnostically; `=0` force-offs everywhere; gfx1151/gfx12 and non-V2
/// gfx1100 stay default-on.
/// default-off graph capability quarantine for gfx1100 + V2 and the measured
/// gfx1201 legacy MQ4 workload; direct batched HIP remains on.
/// `HIPFIRE_VERIFY_GRAPH=1` opts back in diagnostically; `=0` force-offs
/// everywhere; other gfx12 shapes stay default-on.
fn dflash_verify_graph_env_eligible(
arch: &str,
output_dtype: rdna_compute::DType,
Expand All @@ -231,6 +231,11 @@ fn dflash_verify_graph_env_eligible(
if arch == "gfx1100" && is_mq_v2 {
return env_value == Some("1");
}
if arch == "gfx1201" && output_dtype == rdna_compute::DType::MQ4G256 {
// Paired 11.5K DFlash runs on R9700 show graph replay slower than the
// direct tiled-flash path; keep it available as an explicit diagnostic.
return env_value == Some("1");
}
true
}

Expand Down Expand Up @@ -7632,6 +7637,16 @@ mod tests {
DType::MQ3G256V2,
None
));
assert!(!dflash_verify_graph_env_eligible(
"gfx1201",
DType::MQ4G256,
None
));
assert!(dflash_verify_graph_env_eligible(
"gfx1201",
DType::MQ4G256,
Some("1")
));
}

fn try_gpu() -> Option<Gpu> {
Expand Down
Loading