compiled artefacts from win_x64 (sm_120)
Browse files- README.md +12 -0
- code-daemon-embed-v1-l_linux_x64_trt11.3_sm_120.engine +2 -2
- code-daemon-embed-v1-l_ov2026.4_cpu_int8_b64_s80.bin +1 -1
- code-daemon-embed-v1-l_ov2026.4_cpu_int8_b64_s80.xml +0 -0
- code-daemon-embed-v1-l_ov2026.4_igpu_lnl_int8_b64_s80.bin +1 -1
- code-daemon-embed-v1-l_ov2026.4_npu_int4_b16_s80.bin +2 -2
- code-daemon-embed-v1-l_ov2026.4_npu_int4_b16_s80.xml +0 -0
- code-daemon-embed-v1-l_win_x64_trt11.3_sm_120.engine +1 -1
- code-daemon-embed-v1-m_linux_x64_trt11.3_sm_120.engine +2 -2
- code-daemon-embed-v1-m_ov2026.4_cpu_int8_b64_s64.bin +1 -1
- code-daemon-embed-v1-m_ov2026.4_cpu_int8_b64_s64.xml +0 -0
- code-daemon-embed-v1-m_ov2026.4_igpu_lnl_int8_b64_s64.bin +1 -1
- code-daemon-embed-v1-m_ov2026.4_npu_int4_b16_s64.bin +2 -2
- code-daemon-embed-v1-m_ov2026.4_npu_int4_b16_s64.xml +0 -0
- code-daemon-embed-v1-m_win_x64_trt11.3_sm_120.engine +2 -2
- code-daemon-embed-v1-s_linux_x64_trt11.3_sm_120.engine +2 -2
- code-daemon-embed-v1-s_ov2026.4_cpu_int8_b64_s48.bin +1 -1
- code-daemon-embed-v1-s_ov2026.4_cpu_int8_b64_s48.xml +0 -0
- code-daemon-embed-v1-s_ov2026.4_igpu_lnl_int8_b64_s48.bin +1 -1
- code-daemon-embed-v1-s_ov2026.4_npu_int4_b16_s48.bin +2 -2
- code-daemon-embed-v1-s_ov2026.4_npu_int4_b16_s48.xml +0 -0
- code-daemon-embed-v1-s_win_x64_trt11.3_sm_120.engine +2 -2
- code-daemon-embed-v1-xl_linux_x64_trt11.3_sm_120.engine +2 -2
- code-daemon-embed-v1-xl_ov2026.4_cpu_int8_b64_s128.bin +1 -1
- code-daemon-embed-v1-xl_ov2026.4_cpu_int8_b64_s128.xml +0 -0
- code-daemon-embed-v1-xl_ov2026.4_igpu_lnl_int8_b64_s128.bin +1 -1
- code-daemon-embed-v1-xl_ov2026.4_npu_int4_b16_s128.bin +2 -2
- code-daemon-embed-v1-xl_ov2026.4_npu_int4_b16_s128.xml +0 -0
- code-daemon-embed-v1-xl_win_x64_trt11.3_sm_120.engine +2 -2
- manifest.json +2 -2
README.md
CHANGED
|
@@ -17,6 +17,7 @@ pipeline_tag: feature-extraction
|
|
| 17 |
base_model:
|
| 18 |
- intfloat/multilingual-e5-base
|
| 19 |
- Alibaba-NLP/gte-modernbert-base
|
|
|
|
| 20 |
- Qwen/Qwen3-Reranker-4B
|
| 21 |
---
|
| 22 |
|
|
@@ -60,6 +61,12 @@ Laptop RTX 5060, 2026-09-11.
|
|
| 60 |
accuracy is.
|
| 61 |
- **Inside the hybrid search it was built for** (80 real agent queries, file level, runtime
|
| 62 |
reranker off): the right file is first **46 %** of the time and in the top 5 **78 %**.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
- **Its own depth trade:** against its 6-layer parent it gives up ~4 % relative hit@1 for ~1.5×
|
| 64 |
the engine speed.
|
| 65 |
- **Not multilingual.** The vocabulary prune kept 942 of XLM-R's 31 671 Cyrillic pieces; for
|
|
@@ -178,6 +185,11 @@ trtexec --onnx=model_int8qdt.onnx --saveEngine=embed-m.engine --builderOptimizat
|
|
| 178 |
- **Two teachers:** `gte-modernbert-base` for the vector space, and `Qwen3-Reranker-4B` for the
|
| 179 |
ranking — the student learned the cross-encoder's score distribution over hard negatives, not
|
| 180 |
only "positive above negative".
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 181 |
- **INT8 by quantization-aware training**; the E5 prefixes, the pooler and MLM heads and the
|
| 182 |
token-type input were dropped; pooling and L2 were fused into the export.
|
| 183 |
|
|
|
|
| 17 |
base_model:
|
| 18 |
- intfloat/multilingual-e5-base
|
| 19 |
- Alibaba-NLP/gte-modernbert-base
|
| 20 |
+
- google/embeddinggemma-2
|
| 21 |
- Qwen/Qwen3-Reranker-4B
|
| 22 |
---
|
| 23 |
|
|
|
|
| 61 |
accuracy is.
|
| 62 |
- **Inside the hybrid search it was built for** (80 real agent queries, file level, runtime
|
| 63 |
reranker off): the right file is first **46 %** of the time and in the top 5 **78 %**.
|
| 64 |
+
- **October 2026 re-distillation, same body and speed.** The whole training was re-run on a
|
| 65 |
+
1.55M-text serve-format corpus from 173 repositories. On a 10-repository holdout the re-distilled
|
| 66 |
+
weights rank a paraphrase-style query's file in the top 5 **56.5 %** of the time against
|
| 67 |
+
**54.2 %** for the weights this table describes (paraphrase hit@1 34.6 % vs 32.2 %); identifier-style
|
| 68 |
+
queries give up 1.3 points at hit@1 and stay flat at hit@5 (94.3 % vs 94.6 %). Engine throughput is
|
| 69 |
+
unchanged: 14 369 vs 14 606 texts/s, m bucket, INT8, batch 128×64, RTX 5060 — within noise.
|
| 70 |
- **Its own depth trade:** against its 6-layer parent it gives up ~4 % relative hit@1 for ~1.5×
|
| 71 |
the engine speed.
|
| 72 |
- **Not multilingual.** The vocabulary prune kept 942 of XLM-R's 31 671 Cyrillic pieces; for
|
|
|
|
| 185 |
- **Two teachers:** `gte-modernbert-base` for the vector space, and `Qwen3-Reranker-4B` for the
|
| 186 |
ranking — the student learned the cross-encoder's score distribution over hard negatives, not
|
| 187 |
only "positive above negative".
|
| 188 |
+
- **Re-distilled 2026-10-08** on the serving text format itself: 1 546 752 entity texts and queries
|
| 189 |
+
from 173 repositories, built the way the daemon writes them. The dense target is a mix of
|
| 190 |
+
`gte-modernbert-base` and the text tower of `google/embeddinggemma-2` (768-d, bf16), with the
|
| 191 |
+
embeddinggemma vectors rotated onto the student's own coordinates by orthogonal Procrustes first —
|
| 192 |
+
a per-coordinate MSE pays for the rotation out of the training budget, and the rotation is free.
|
| 193 |
- **INT8 by quantization-aware training**; the E5 prefixes, the pooler and MLM heads and the
|
| 194 |
token-type input were dropped; pooling and L2 were fused into the export.
|
| 195 |
|
code-daemon-embed-v1-l_linux_x64_trt11.3_sm_120.engine
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:086188e73351129e0e5a066041ae59b8c72d0c1878e1685f4e9c4a57969b3388
|
| 3 |
+
size 101787076
|
code-daemon-embed-v1-l_ov2026.4_cpu_int8_b64_s80.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 100617601
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f07a06f83d610359e510ab80a2f68b6211c0393614437df52552a47b8f7f3df6
|
| 3 |
size 100617601
|
code-daemon-embed-v1-l_ov2026.4_cpu_int8_b64_s80.xml
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
code-daemon-embed-v1-l_ov2026.4_igpu_lnl_int8_b64_s80.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 100617601
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f07a06f83d610359e510ab80a2f68b6211c0393614437df52552a47b8f7f3df6
|
| 3 |
size 100617601
|
code-daemon-embed-v1-l_ov2026.4_npu_int4_b16_s80.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8ec2c4a8bb009aa9dd47ab29f276e62e55e49068fb9fb97e48b92293fafc3acd
|
| 3 |
+
size 33285222
|
code-daemon-embed-v1-l_ov2026.4_npu_int4_b16_s80.xml
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
code-daemon-embed-v1-l_win_x64_trt11.3_sm_120.engine
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 101786564
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f75df3f83259f007b066bf85ffb026a258ec84c52988e873f998603c578459e7
|
| 3 |
size 101786564
|
code-daemon-embed-v1-m_linux_x64_trt11.3_sm_120.engine
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:386e6197219b3efe3e77726ca49f04c7c1adceea3b1e1198f05b12e8903e70e2
|
| 3 |
+
size 101771580
|
code-daemon-embed-v1-m_ov2026.4_cpu_int8_b64_s64.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 100617601
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f07a06f83d610359e510ab80a2f68b6211c0393614437df52552a47b8f7f3df6
|
| 3 |
size 100617601
|
code-daemon-embed-v1-m_ov2026.4_cpu_int8_b64_s64.xml
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
code-daemon-embed-v1-m_ov2026.4_igpu_lnl_int8_b64_s64.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 100617601
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f07a06f83d610359e510ab80a2f68b6211c0393614437df52552a47b8f7f3df6
|
| 3 |
size 100617601
|
code-daemon-embed-v1-m_ov2026.4_npu_int4_b16_s64.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8ec2c4a8bb009aa9dd47ab29f276e62e55e49068fb9fb97e48b92293fafc3acd
|
| 3 |
+
size 33285222
|
code-daemon-embed-v1-m_ov2026.4_npu_int4_b16_s64.xml
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
code-daemon-embed-v1-m_win_x64_trt11.3_sm_120.engine
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ea7c32a0a19ec49d7bf3b161d1452f84dc13d11127c0e91da1b973a7bd5cd1bc
|
| 3 |
+
size 101785964
|
code-daemon-embed-v1-s_linux_x64_trt11.3_sm_120.engine
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c91b58c6111301c6f834e4cd6c75cb4490a53faf10e42be3e1a6f596fd2bd354
|
| 3 |
+
size 101861348
|
code-daemon-embed-v1-s_ov2026.4_cpu_int8_b64_s48.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 100617601
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f07a06f83d610359e510ab80a2f68b6211c0393614437df52552a47b8f7f3df6
|
| 3 |
size 100617601
|
code-daemon-embed-v1-s_ov2026.4_cpu_int8_b64_s48.xml
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
code-daemon-embed-v1-s_ov2026.4_igpu_lnl_int8_b64_s48.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 100617601
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f07a06f83d610359e510ab80a2f68b6211c0393614437df52552a47b8f7f3df6
|
| 3 |
size 100617601
|
code-daemon-embed-v1-s_ov2026.4_npu_int4_b16_s48.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8ec2c4a8bb009aa9dd47ab29f276e62e55e49068fb9fb97e48b92293fafc3acd
|
| 3 |
+
size 33285222
|
code-daemon-embed-v1-s_ov2026.4_npu_int4_b16_s48.xml
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
code-daemon-embed-v1-s_win_x64_trt11.3_sm_120.engine
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b688c6b7b247686f3ad3f2170c7251158c6ca906c3ceea537a3d6ffd0812d18c
|
| 3 |
+
size 101847436
|
code-daemon-embed-v1-xl_linux_x64_trt11.3_sm_120.engine
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:893ce104c83eba45b363755d0815a5d18b086dc95d6c02bd1342df0f5e09f3f9
|
| 3 |
+
size 101789916
|
code-daemon-embed-v1-xl_ov2026.4_cpu_int8_b64_s128.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 100617601
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f07a06f83d610359e510ab80a2f68b6211c0393614437df52552a47b8f7f3df6
|
| 3 |
size 100617601
|
code-daemon-embed-v1-xl_ov2026.4_cpu_int8_b64_s128.xml
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
code-daemon-embed-v1-xl_ov2026.4_igpu_lnl_int8_b64_s128.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 100617601
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f07a06f83d610359e510ab80a2f68b6211c0393614437df52552a47b8f7f3df6
|
| 3 |
size 100617601
|
code-daemon-embed-v1-xl_ov2026.4_npu_int4_b16_s128.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8ec2c4a8bb009aa9dd47ab29f276e62e55e49068fb9fb97e48b92293fafc3acd
|
| 3 |
+
size 33285222
|
code-daemon-embed-v1-xl_ov2026.4_npu_int4_b16_s128.xml
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
code-daemon-embed-v1-xl_win_x64_trt11.3_sm_120.engine
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:25e138d182efc6262ac52a2cce0749f23da129e16048482cc7e436c77ef0d824
|
| 3 |
+
size 101787340
|
manifest.json
CHANGED
|
@@ -5,7 +5,7 @@
|
|
| 5 |
"quantization": "int8",
|
| 6 |
"targets": [
|
| 7 |
"openvino",
|
| 8 |
-
"
|
| 9 |
],
|
| 10 |
"runtime_tags": {
|
| 11 |
"trt": "trt11.3",
|
|
@@ -13,6 +13,6 @@
|
|
| 13 |
"tvm": "tvm0.25",
|
| 14 |
"mlx_dir": "model_gpu_mlx0.32"
|
| 15 |
},
|
| 16 |
-
"compiled_at": "2026-
|
| 17 |
"compiled_by": "models/_compile"
|
| 18 |
}
|
|
|
|
| 5 |
"quantization": "int8",
|
| 6 |
"targets": [
|
| 7 |
"openvino",
|
| 8 |
+
"tensorrt"
|
| 9 |
],
|
| 10 |
"runtime_tags": {
|
| 11 |
"trt": "trt11.3",
|
|
|
|
| 13 |
"tvm": "tvm0.25",
|
| 14 |
"mlx_dir": "model_gpu_mlx0.32"
|
| 15 |
},
|
| 16 |
+
"compiled_at": "2026-10-08T18:07:19Z",
|
| 17 |
"compiled_by": "models/_compile"
|
| 18 |
}
|