Zhongzhu commited on
Commit
6cadfbe
·
verified ·
1 Parent(s): b645aa3

Add GLM-5.2-FP8 MLA latent c_kv rotation (pure Rcov·P·Hblock, 78 layers, g128)

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. GLM-5.2-FP8/c_kv_rcov_phblock_g128/README.md +16 -0
  2. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_0.pt +3 -0
  3. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_1.pt +3 -0
  4. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_10.pt +3 -0
  5. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_11.pt +3 -0
  6. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_12.pt +3 -0
  7. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_13.pt +3 -0
  8. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_14.pt +3 -0
  9. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_15.pt +3 -0
  10. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_16.pt +3 -0
  11. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_17.pt +3 -0
  12. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_18.pt +3 -0
  13. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_19.pt +3 -0
  14. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_2.pt +3 -0
  15. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_20.pt +3 -0
  16. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_21.pt +3 -0
  17. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_22.pt +3 -0
  18. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_23.pt +3 -0
  19. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_24.pt +3 -0
  20. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_25.pt +3 -0
  21. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_26.pt +3 -0
  22. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_27.pt +3 -0
  23. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_28.pt +3 -0
  24. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_29.pt +3 -0
  25. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_3.pt +3 -0
  26. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_30.pt +3 -0
  27. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_31.pt +3 -0
  28. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_32.pt +3 -0
  29. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_33.pt +3 -0
  30. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_34.pt +3 -0
  31. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_35.pt +3 -0
  32. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_36.pt +3 -0
  33. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_37.pt +3 -0
  34. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_38.pt +3 -0
  35. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_39.pt +3 -0
  36. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_4.pt +3 -0
  37. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_40.pt +3 -0
  38. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_41.pt +3 -0
  39. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_42.pt +3 -0
  40. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_43.pt +3 -0
  41. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_44.pt +3 -0
  42. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_45.pt +3 -0
  43. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_46.pt +3 -0
  44. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_47.pt +3 -0
  45. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_48.pt +3 -0
  46. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_49.pt +3 -0
  47. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_5.pt +3 -0
  48. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_50.pt +3 -0
  49. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_51.pt +3 -0
  50. GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_52.pt +3 -0
GLM-5.2-FP8/c_kv_rcov_phblock_g128/README.md ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # GLM-5.2-FP8 — MLA latent `c_kv` rotation (pure Rcov·P·Hblock)
2
+
3
+ Per-layer orthogonal rotations for INT2 quantization of the MLA shared latent
4
+ `c_kv` (kv_lora_rank=512), all 78 layers. Composition: covariance eigenvectors
5
+ → bit-reversal permutation → per-group block-diag-128 Hadamard, fitted on a
6
+ GLM-5.2-FP8 `c_kv` calibration dump (GPQA prompts). `k_pe` (RoPE, 64d) stays
7
+ BF16. Files: `layer_<i>.pt`, each a `[512, 512]` float32 orthogonal matrix
8
+ (ortho_err ~4e-7).
9
+
10
+ Serve with the OSCAR SGLang fork:
11
+ ```
12
+ SGLANG_OSCAR_MLA_KV_ROTATION_PATH=<this dir> \
13
+ SGLANG_OSCAR_MLA_KV_GROUP_SIZE=128 \
14
+ SGLANG_LLOYD_MAX=1 \
15
+ python -m sglang.launch_server --model-path zai-org/GLM-5.2-FP8 ...
16
+ ```
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_0.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:04cffda83c97d8cf661f072638562128c2e7c1a87ac906b59ba957d774ffbbcc
3
+ size 1050153
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_1.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0779116f9444d7056f0339d2775dc9b3b9229ec18000444f7e44cd8c2a259ae5
3
+ size 1050153
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_10.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc66b6749b90a3b1ffb059dd1b7f0351a7ac56b801923ed6d14cc271b7602812
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_11.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:14851261ac8273abb363d273a5e451ace31df3e4d292f9a6b249ab518ff431c8
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_12.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:abed7bb9ebe7ea99598b2799a71da6aaa18a7938f435e49030419e95f8447c1f
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_13.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a5e5466b1ff581e688af0b80da7eab3f2ca3b8753b3780c0cef82062c80712c0
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_14.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1ae058220dd721e6b0402143a16a9e035c3e06728c411a552e7de20f964a197a
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_15.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e50bcd36dd73d7af43bfef7b5c36cfa51c02a41cd65d4ca39271b375ce9bb857
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_16.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:254817477c105feedac8644a735fa31852811556017a2601121a95f9116c1d30
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_17.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:106f70303aab207b977d9b0b97b5bfd3ae891f28087fade789e4c06628d5d4a0
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_18.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0a74b310c87584dc9ee12cfef649f6f43ead2375702e4eef99be3e3f4390a086
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_19.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8acdc98579ee03a489fbd5f605ba30e344795a564596794ff1646b42996ea138
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_2.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:11e0dcef3b71af310ac5c182caddccf75f2f79877a9ec4f3ead16fa785f06a76
3
+ size 1050153
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_20.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:98d2a4eb72bb08eea0695eaf8c7545add1dfb2b3009ac664ab377525834b4434
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_21.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:044c2908549b40680fd72f589db5b83f367f1cd711f84c420ed0b99bcd1fcdbf
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_22.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:403a9de1cbeb5868dafdff328936f8c2fbc8338f43b5f18d3311f8089828aa31
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_23.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0fd29d78efc094bd9ecd3809904e67ead90ac61eeb6e0e280104035d2a9bce95
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_24.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:961baae6920b8463ea5cb3a35b8446ed4214f2d4e51cb30c8f06fc9d988bc7f1
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_25.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:abaf691ba621e40b0e6ed99d929d554481275f758461c4306dfc7feb0e76bb45
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_26.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2fa1667b2fe3216ccea1e77dbebf744f6f4136795d1bd0dc41fcecd52b95bed
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_27.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1dfd66caf4144b146f361be08bd315eddce4f696836595d62f5d6256e496197d
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_28.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:32b41483f9db5a9e502e218e5e85cd984f31c92472a86aad483ea758226f8514
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_29.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:47ea27cd212f873bd797abc1899ec652e1ef5fd1d0e5490787a9849b34ccf6ac
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_3.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:baad71470b8a7ad64d770aa4af960626fa3d4f0cf139f9dc81f5273d686f1b5a
3
+ size 1050153
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_30.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e5600171828b074523db41d6ef561df00b56a1625d95bbcf483d0d04fd74e3b
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_31.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:56920cd61c0e1d02313e714325ab4f1893108769a44a5887629403652c2ec280
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_32.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f4d8631d768e738a183e2b186ac8cf8cd3ddadc7e39d6be5073fc97a261f5131
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_33.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c306bf08217378dfe88aa30bbb66a179a954718ac3e4ef9d8b960f6287efd745
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_34.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9667303c1d9cb4bc5189fd2511f8cefca82109e9c5a4c0dbd8d1a8e4294f6bca
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_35.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:91366489f15d2734b4c7fcfba7ed06606c98fd68b045eb4140dffeb6a6548661
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_36.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e9071ce56963c792201e7d75d0f5a012aaebbb309d68b757675a2f6b999dcca2
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_37.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:49f06f2f4c4d6aade9d0d03c9560e66f7756e4cbddac0eef819d780d06aabe98
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_38.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5e789b01fa7f033254524e6dddd8e8d688b3c603162394978c5eadda70d38050
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_39.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:474cd05726b26d95b965b53317db43b83716a24418df7e3b4f4bafa08c42a8bc
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_4.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:315b0368edcc203c33b7aceece54c3e67cd14eb10c5b01999f387168c65303e8
3
+ size 1050153
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_40.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0f6234734418fc7383eb89b161c5e20cde842118b0725ae5f68daa75e7bd8bfa
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_41.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5606454a2fe167c41ed42068282277edf766e1cb42364d0db88801532bc68ee5
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_42.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba1ea3fd2b32f83fddb882e868a0875a4aeb49f9f4b5f51fdd46d30567e97d8f
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_43.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05b0cdd6e8e0b0932b9a6514bb83721640fcfdf7dd08397a72c791699736b793
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_44.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4b96cb4b58dbdfe347b1f0319f9aee024235cb792a2caef66021fbfcb703f9ac
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_45.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d3b48486819e322b83e44ebda706c3932e7db03579e51a7db290898d02e8078
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_46.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f1943d5e7e38a94f8a311f15eb769d4d07e1ec63505398c429669e0a323fa4d3
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_47.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:006d8fe8ee5416b03ec802347eb6008fc1d26fa9caae6a0478816bf74ad11fb0
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_48.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:79b78bd7c499d36b3b69d9dc8cf3d0902eef63bfcd59d641af6a907d2b214d09
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_49.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2b197e7c2de1290c1861306bd3337b2d60fba50ef6c8916c16d7c866a296d26d
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_5.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:91b430a25e8abf8fc47ffa30ad7c34ab38a7c4db2297b67db45bae1d4096ce1f
3
+ size 1050153
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_50.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:837822f52b0b0a048f0c91e50b762e7c306ea285c7b9578ac42f692252696a79
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_51.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e283172398e85fab31ded7eeca867ab0fcc0760eb0cb3c7f869ad296d56671c
3
+ size 1050160
GLM-5.2-FP8/c_kv_rcov_phblock_g128/layer_52.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ea7cd5b95918ed56dae3bc8f95c4b4692e8fec7bc285793cbd2b159bedad7a11
3
+ size 1050160