glm cuda: prompt chunks of 4,096 rows, and the DFlash2 context buffers sized to match

zig/src/families/glm/cuda_state.zig

@@ -16,7 +16,7 @@ const mlp = @import("cuda_mlp.zig");
 const kda_block = @import("cuda_kda_block.zig");
 
 /// A prompt chunk's rows (decode.PREFILL_ROWS).
-pub const prefill_rows = 2048;
+pub const prefill_rows = 4096;
 /// A decode window's rows (Engine's max_rows).
 pub const window_rows = 8;
 /// Prompts one prompt pass takes together (cuda_prompts): a prompt buffer's logits rows, one a prompt.

zig/src/families/glm/dflash.zig

@@ -27,7 +27,7 @@ pub const CH: usize = 16; // key chunks of the chunked attention (df_attn_part)
 pub const CW: usize = 2 * TOPK; // a row's gathered ints: 16 ids, then 16 values' bits
 pub const MAX_S: usize = 16; // streams one block pass drafts for
 pub const MAX_M: usize = MAX_S * BLOCK; // block rows at most
-pub const MAX_CTX: usize = 2048; // context rows one ingest takes (a prompt chunk's)
+pub const MAX_CTX: usize = @import("cuda_state.zig").prefill_rows; // context rows one ingest takes (a prompt chunk's)
 pub const STEP_ROWS: usize = 32; // a verify's tap rows (the shared windows' rows at most)
 pub const TAP_ROW: usize = taps_mod.layers.len * D * 2; // bytes of one row's taps