bogoconic1/sft-test-step-1000-opd-deploy-fixed

sft-test-step-1000 OPD deployment repair

This repository preserves the trained weights from jacobmorrison/sft-test-step-1000 and repairs only the deployment metadata. The source checkpoint is an SFT continuation of opd-32b-deploy, but its export omitted or rewrote files required by the OPD SGLang deployment stack.

Unchanged assets

• Every model tensor is byte-for-byte unchanged from jacobmorrison/sft-test-step-1000 at revision 3d1ce735b2c46d1800e38857995818f681ca7265.
• The model contains 32,530,505,216 parameters and 65,061,010,432 bytes of tensor data.
• The 771 weight keys exactly match the opd-32b-deploy weight-key schema, including all 64 attention-sink tensors.
• tokenizer.json and tokenizer_config.json are unchanged. Their source files are byte-identical to the OPD deployment tokenizer.

Repaired deployment files

config.json

Replaced the generic Transformers 5.13 export with the deployment configuration from opd-32b-deploy. The trained architecture and dimensions are unchanged. This restores:

• the flattened YaRN rope_scaling structure required by the deployed SGLang RoPE factory;
• max_position_embeddings: 262144;
• is_hybrid_swa: true and the 64-entry 3:1 sliding/full attention pattern;
• sink_init_value: 0.0;
• reuse_packing_metadata: true; and
• explicit BF16 dtype metadata.

The source SFT export used nested full_attention and sliding_attention RoPE dictionaries. In the deployed SGLang stack that reached the RoPE factory as an unhashable dictionary and prevented the model from loading.

generation_config.json

Replaced the malformed training export with the OPD deployment generation configuration. The source file contained eos_token_id: [null, null] and used pad_token instead of pad_token_id. The repaired file restores token IDs 0/1/2, sampling defaults, and the 32,768-token default generation ceiling.

chat_template.jinja

Added the byte-identical OPD deployment chat template, which was absent from the source SFT repository. This makes OpenAI-compatible chat requests render with the same message format used to evaluate and deploy the base model.

Intended runtime

The repaired bundle is intended for the same SGLang OPD runtime as the base deployment. It supports BF16 target inference and the compatible phase-L DFlash draft without modifying the SFT weights.