Skip to content

glm: fix streaming prefill failures for real-size prompts #520

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
andreaborio wants to merge 2 commits into
base: glm5.2
Choose a base branch
from
Open

glm: fix streaming prefill failures for real-size prompts #520

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ee02641
glm: honor the full-layer prefill threshold in the indexed prefill flow
Jul 8, 2026
b85f6f8
glm: evict at budget in the batch selected-expert prefill loader
Jul 8, 2026
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
12 changes: 10 additions & 2 deletions ds4.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30722,7 +30722,15 @@ static bool glm_graph_forward_indexed_tokens(
const uint32_t drain_interval =
progress_flush_interval != 0 ? glm_graph_indexed_prefill_drain_interval() : 0u;
const bool progress_requested = display_progress && work_total > 0;
ds4_gpu_set_glm_streaming_prefill_full_layer(false);
/* Mirror the plain prefill flow: above the full-layer threshold the
* selected-expert addr loader is the wrong tool — a chunk of N tokens
* selects far more unique experts per layer than the cache can hold, and
* its overflow branch needs the full expert tensors mapped. Hardcoding
* false here broke every generation prompt above ~64 tokens ("failed to
* map overflow expert views"), while short prompts keep the addr path. */
const bool full_layer_prefill =
glm_graph_stream_prefill_full_layer_enabled(g, n_tokens);
ds4_gpu_set_glm_streaming_prefill_full_layer(full_layer_prefill);

if (trace) {
glm_graph_indexed_prefill_tracef(
Expand Down Expand Up @@ -30818,7 +30826,7 @@ static bool glm_graph_forward_indexed_tokens(
weights,
il,
n_tokens,
false);
full_layer_prefill);
if (ok) ok = ds4_gpu_begin_commands() != 0;
}
const ds4_layer_weights *l = &weights->layer[il];
Expand Down
16 changes: 10 additions & 6 deletions ds4_metal.m
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12859,12 +12859,16 @@ static int ds4_gpu_stream_expert_cache_prepare_selected_batch(
continue;
}

if (cache_budget != 0 && reserved_entries >= cache_budget) {
unique_entries[u] = NULL;
continue;
}

const int force_reuse = 0;
/* At budget, evict-and-reuse like the decode loader does instead
* of spilling to whole-tensor overflow views: the overflow wrap
* needs the expert tensors in the mapped view set, which the
* decode-expert-cache prefill mapping does not provide, so a
* full cache made any checkpoint-extend or small-suffix prefill
* fail ("failed to map overflow expert views"). Commands were
* just synced by the caller, so reuse cannot race the GPU, and
* prepare_load_buffers protects this layer's unique set. */
const int force_reuse =
cache_budget != 0 && reserved_entries >= cache_budget;
ds4_gpu_stream_expert_readahead_range(unique_gate_offsets[u],
gate_expert_bytes);
ds4_gpu_stream_expert_readahead_range(unique_up_offsets[u],
Expand Down