metal : key the fa-vec tuned table by family instead of SKU (#29075)

* key the fa-vec tuned table by family instead of SKU

* fall back to baseline for untuned fa-vec gpu families

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -3584,7 +3584,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         auto cfg = use_sparse
                 ? ggml_metal_tuning::fa_vec_baseline_cfg((int) ne00, (int) ne20)
                 : ggml_metal_tuning::fa_vec_pick(
-                          props_dev->device_id,
                           props_dev->gpu_family,
                           (int) op->src[1]->type,
                           (int) ne00, (int) ne20,   // dk, dv (ne00 == dk for FA)

ggml/src/ggml-metal/ggml-metal-tuning.h

@@ -1,6 +1,5 @@
 #pragma once
 
-#include "ggml-metal-device.h"  // enum ggml_metal_device_id
 #include "ggml.h"
 
 #include <cstdint>
@@ -32,7 +31,7 @@ constexpr int8_t FA_VEC_DOMAIN_DECODE = 0;  // ne01 == 1
 constexpr int8_t FA_VEC_DOMAIN_BATCH  = 1;  // ne01 >= 2
 
 struct fa_vec_key_t {
-    int8_t  device_id;
+    int8_t  family;
     int8_t  dtype;
     int16_t dk;
     int16_t dv;
@@ -70,8 +69,7 @@ void         fa_vec_set_override(fa_vec_cfg_t cfg);
 void         fa_vec_clear_override();
 fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv);
 
-// device_id selects a per-SKU row; on a miss, gpu_family (0 if unknown) maps to a representative
-// SKU and the table is retried. No match -> baseline.
-fa_vec_cfg_t fa_vec_pick(enum ggml_metal_device_id device_id, int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01);
+// Keyed by Apple GPU family; an untuned family matches no row and gets the baseline.
+fa_vec_cfg_t fa_vec_pick(int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01);
 
 }  // namespace ggml_metal_tuning

tools/tuning/README.md

@@ -26,10 +26,13 @@ Sweep the grid (6 dtypes x 10 head sizes x 4 KV depths x 9 batch widths; a few h
 ./build/bin/ggml-metal-tuning fa-vec > fa_vec_rows.txt 2> fa_vec_sweep.log
 ```
 
-`fa_vec_rows.txt` holds nothing but table rows, ready to paste into `fa_vec_tuned_table`: the min-max-regret target, the aggregate benefit gate, the short-KV drop and the pointwise compression are already applied.
+`fa_vec_rows.txt` holds nothing but table rows: the min-max-regret target, the aggregate benefit gate, the short-KV drop and the pointwise compression are already applied.
+The rows carry the SKU token the runtime reported, but `fa_vec_tuned_table` is keyed by Apple GPU family, so that column has to be retagged before the rows compile.
+Your family number is on the `MTLGPUFamilyApple<N>` line the backend logs at init, near the top of `fa_vec_sweep.log`; `N` is the value, and the `MTLGPUFamilyCommon`/`MTLGPUFamilyMetal` lines beside it are not it.
+If your log is the only sweep for that family, its rows become the family's segment; where the family already has rows, post the log and let the two be compared before anything is replaced.
 A config represents a bucket only if it is no slower than the baseline config at every point that bucket covers, so a config that wins on average but loses at one batch width leaves its bucket at baseline.
 `fa_vec_sweep.log` holds the per-cell timings, bucket coverage, noise floor, any cooldown activity, and every config the no-harm rule refused together with the point that refused it.
-Post both: the log is what makes the rows reviewable.
+Post both, always: the rows now speak for every device in the family, so the log is what makes them reviewable.
 
 Long sweeps can be split.
 `--dtype f16,q4_0` and `--dk 128,192` restrict the grid, and the emitted rows for one `(dtype, head size)` do not depend on the others.