chenjiel commited on
Commit
b6220d6
·
unverified ·
1 Parent(s): eb4ad80

Support BLASST sparsity

Browse files
Files changed (1) hide show
  1. config.json +27 -2
config.json CHANGED
@@ -264,5 +264,30 @@
264
  "use_mtp": true,
265
  "use_qk_norm": true,
266
  "use_routing_bias": true,
267
- "vocab_size": 200064
268
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
264
  "use_mtp": true,
265
  "use_qk_norm": true,
266
  "use_routing_bias": true,
267
+ "vocab_size": 200064,
268
+ "sparse_attention_config": {
269
+ "config_groups": {
270
+ "group_0": {
271
+ "sparse_algo": "softmax_skip",
272
+ "targets": [
273
+ "MiniMaxM2Attention"
274
+ ]
275
+ }
276
+ },
277
+ "threshold_scale_factor": {
278
+ "formula": "a * exp(b * target_sparsity)",
279
+ "prefill": {
280
+ "a": 87.239981,
281
+ "b": 5.3221
282
+ },
283
+ "decode": {
284
+ "a": 0.032140,
285
+ "b": 11.7877
286
+ }
287
+ },
288
+ "producer": {
289
+ "name": "modelopt",
290
+ "version": "0.43.0.dev27+g26aa1d0b1"
291
+ }
292
+ }
293
+ }