[ET-VK][sdpa] Reuse shared V cache across GQA query heads in AV coop-GEMV#21117
Merged
Conversation
…GEMV Pull Request resolved: #21063 The LLM decode AV coop-GEMV reloads the shared V cache once per query head. In grouped-query attention Hq = G * Hkv query heads share each KV head (Llama G=4, Phi G=3, Qwen G=2), and out[q_h, d] = sum_c attn[c, q_h] * V[c, kv_h, d] reads the SAME V texel for every query head in a group. The per-query-head coop shader gives each of the Hq heads its own workgroup, so V -- the dominant traffic (head_dim-wide per context texel, vs a scalar attn weight) -- is read G times. This adds a GQA-reuse AV variant that assigns ONE workgroup per (d4, kv_h): it loads each V texel once and reuses it across all G query heads in the group, producing G output texels. For this bandwidth-bound kernel that cuts V-cache traffic ~Gx. Implementation: - The variant is a codegen flag (`GQA`) on the existing `sdpa_compute_out_coop.glsl` template, not a separate file: one shared header plus two `#ifdef GQA` `main()`s (per-head and GQA-reuse), so the shared setup lives in one place while each algorithm reads end-to-end. It emits the shader `sdpa_compute_out_gqa_coop`. - Reduction reuses the per-head coop shader's shared-memory tree reduction (no subgroup arithmetic), so the variant runs on any Vulkan device -- Adreno and Mali alike -- with no capability gate. - Each thread holds G output accumulators; the array is sized to a compile-time `MAX_GROUP_SIZE` = 8 and the group loop is bounded by the `group_size` = Hq/Hkv spec constant, so the driver fully unrolls it at pipeline creation. - Dispatch (`pick_sdpa_av_shader` + global-wg picker + spec-const wiring in `add_sdpa_compute_out_node`): the GQA variant is selected on the LLM decode coop path when Hq > Hkv, evenly divisible, and G <= 8 (`use_gqa_av_coop`); it sets `group_size` and changes the global workgroup z-dim from Hq to Hkv. Everything else -- MHA (Hq == Hkv), groups exceeding the cap (G > 8, e.g. MQA with Hq > 8), and non-divisible shapes -- falls back to the unchanged per-head `sdpa_compute_out_coop`. (Low-ratio MQA -- Hkv == 1 with Hq <= 8 -- is eligible and takes the GQA path.) - A test-only `gqa_override` knob is threaded through `add_sdpa_compute_out_node` (declared in the new `SDPA.h`): -1 auto-select, 0 force per-head, 1 force GQA, so a benchmark can exercise both AV shaders on the same shape; forcing GQA is VK_CHECK'd against shape eligibility. ghstack-source-id: 405400503 @exported-using-ghexport Differential Revision: [D112906311](https://our.internmc.facebook.com/intern/diff/D112906311/)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21117
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
…QA AV coop-GEMV Pull Request resolved: #21064 Builds on the GQA-reuse AV coop-GEMV (parent commit). The base `sdpa_compute_out_gqa_coop` shader assigns one head_dim texel (TILE_N4=1) per workgroup x-slot. This adds a head_dim output-tiled variant, `sdpa_compute_out_gqa_coop_tile2` (TILE_N4=2): each workgroup owns 2 head_dim texels and emits G x TILE_N4 outputs, so the per-context-texel attn-weight loads and the shared-memory tree reduction are amortized over twice as many outputs. The AV coop GEMV GLSL is already a TILE_N4-parameterized template (the `partial_n_tile` uniform branch handles the D4 % TILE_N4 != 0 tail), so this change is purely a new codegen variant plus dispatch wiring — no shader-body change. Selection is vendor-adaptive: `pick_sdpa_av_shader` appends the `_tile2` suffix only on Adreno (`graph->device_is_adreno()`). Tiling is a consistent win on Adreno (AV ~1.14-1.63x over the base GQA variant) but a regression on Mali at common decode contexts (~0.67-0.86x, interleaved median-of-N), so Mali and other vendors keep the base `sdpa_compute_out_gqa_coop`. `pick_sdpa_av_global_wg_size` keys off the same `_tile2` suffix: the tiled variant's x-dim collapses from D4 to div_up(D4, 2) since each workgroup now covers 2 head_dim texels. Because the variant is Adreno-only in production, the test-only `gqa_override` knob (see SDPA.h) is extended so tests can pin the variant on any device: `kGqaOverrideForceTile2` / `kGqaOverrideForceBase` force the tiled / base variant regardless of vendor (via `resolve_use_tile2`), giving the tiled shader deterministic coverage on Mali / SwiftShader. `resolve_use_gqa` is unchanged (any non-`ForceNonGqa` value still forces the GQA family, VK_CHECK'd for eligibility). ghstack-source-id: 405400514 @exported-using-ghexport Differential Revision: [D112906313](https://our.internmc.facebook.com/intern/diff/D112906313/)
SS-JIA
approved these changes
Jul 22, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #21063 by @SS-JIA
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/576/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/576/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/575/orig
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/576/orig
Differential Revision: D112906311
@diff-train-skip-merge