clocksmith commited on
Commit
1c509e3
·
verified ·
1 Parent(s): b5d327d

Publish qwen-3-5-0-8b-q4k-ehaf16 manifest update

Browse files
models/qwen-3-5-0-8b-q4k-ehaf16/manifest.json CHANGED
@@ -235,6 +235,11 @@
235
  "entry": "main",
236
  "digest": "sha256:142817adcacb28750a92793996cd15c2d9fe13d03e0c2d0aadf3e674dc7a35d6"
237
  },
 
 
 
 
 
238
  "tiled": {
239
  "kernel": "matmul_f16w_f32a.wgsl",
240
  "entry": "main",
@@ -248,6 +253,14 @@
248
  "kvDtype": "f16"
249
  }
250
  },
 
 
 
 
 
 
 
 
251
  "lm_head_gemv": {
252
  "kernel": "matmul_gemv_subgroup_f16a.wgsl",
253
  "entry": "main_multicol",
@@ -360,17 +373,17 @@
360
  ],
361
  [
362
  "q_proj",
363
- "q4_prefill",
364
  "layer.{L}.self_attn.q_proj"
365
  ],
366
  [
367
  "k_proj",
368
- "q4_prefill",
369
  "layer.{L}.self_attn.k_proj"
370
  ],
371
  [
372
  "v_proj",
373
- "q4_prefill",
374
  "layer.{L}.self_attn.v_proj"
375
  ],
376
  [
@@ -383,11 +396,11 @@
383
  ],
384
  [
385
  "attention",
386
- "attn_stream"
387
  ],
388
  [
389
  "o_proj",
390
- "q4_prefill",
391
  "layer.{L}.self_attn.o_proj"
392
  ],
393
  [
@@ -400,12 +413,12 @@
400
  ],
401
  [
402
  "gate_proj",
403
- "q4_prefill",
404
  "layer.{L}.mlp.gate_proj"
405
  ],
406
  [
407
  "up_proj",
408
- "q4_prefill",
409
  "layer.{L}.mlp.up_proj"
410
  ],
411
  [
@@ -414,7 +427,7 @@
414
  ],
415
  [
416
  "down_proj",
417
- "q4_prefill",
418
  "layer.{L}.mlp.down_proj"
419
  ],
420
  [
 
235
  "entry": "main",
236
  "digest": "sha256:142817adcacb28750a92793996cd15c2d9fe13d03e0c2d0aadf3e674dc7a35d6"
237
  },
238
+ "q4_widetile": {
239
+ "kernel": "fused_matmul_q4_widetile.wgsl",
240
+ "entry": "main",
241
+ "digest": "sha256:5631ce6f19f805c7ac85722bac17365048bc7b9e2c422f3ef9477b63cb8dad62"
242
+ },
243
  "tiled": {
244
  "kernel": "matmul_f16w_f32a.wgsl",
245
  "entry": "main",
 
253
  "kvDtype": "f16"
254
  }
255
  },
256
+ "attn_head256": {
257
+ "kernel": "attention_head256_f16kv.wgsl",
258
+ "entry": "main",
259
+ "digest": "sha256:4ecdc079e322a770350414244cb383d72ae7ffa47ed620c64ff425c93413e97a",
260
+ "precision": {
261
+ "kvDtype": "f16"
262
+ }
263
+ },
264
  "lm_head_gemv": {
265
  "kernel": "matmul_gemv_subgroup_f16a.wgsl",
266
  "entry": "main_multicol",
 
373
  ],
374
  [
375
  "q_proj",
376
+ "q4_widetile",
377
  "layer.{L}.self_attn.q_proj"
378
  ],
379
  [
380
  "k_proj",
381
+ "q4_widetile",
382
  "layer.{L}.self_attn.k_proj"
383
  ],
384
  [
385
  "v_proj",
386
+ "q4_widetile",
387
  "layer.{L}.self_attn.v_proj"
388
  ],
389
  [
 
396
  ],
397
  [
398
  "attention",
399
+ "attn_head256"
400
  ],
401
  [
402
  "o_proj",
403
+ "q4_widetile",
404
  "layer.{L}.self_attn.o_proj"
405
  ],
406
  [
 
413
  ],
414
  [
415
  "gate_proj",
416
+ "q4_widetile",
417
  "layer.{L}.mlp.gate_proj"
418
  ],
419
  [
420
  "up_proj",
421
+ "q4_widetile",
422
  "layer.{L}.mlp.up_proj"
423
  ],
424
  [
 
427
  ],
428
  [
429
  "down_proj",
430
+ "q4_widetile",
431
  "layer.{L}.mlp.down_proj"
432
  ],
433
  [