mistral-7b-instruct-v0.3-DynQuant-4bit

Mistral-7B-Instruct-v0.3 text-to-SQL, DynQuant 4-bit

mistralai/Mistral-7B-Instruct-v0.3 fine-tuned on text-to-SQL and quantized to 4 bits with DynQuant. It is one of 9 arms in a panel where every quantized arm was allocated the same byte budget, so the accuracies below differ by method and not by size.

What this is

Results

Execution match on 2,454 held-out text-to-SQL problems: the generated query is run against the schema and compared to the reference result set.

This arm, by evaluation source:

How this arm compares

McNemar exact over the per-item hits, so every row is a paired test on the same problems in the same order. p (Holm) is step-down corrected within the family the panel declared, not within this card.

What is not claimed

• The accuracy above was measured in bf16, not from this directory. A DynQuant arm is scored by encoding its allocated widths back into bf16 -- the same encoder, the same widths, the same values -- so that every arm in the panel is scored through one path and no arm's number depends on which container it was read from. The directory you are downloading holds those same values packed. What is carried across from the measurement is the arithmetic; what is not is a claim that the packed and encoded containers were separately scored.
• The arms above are not all the same scheme. In this panel GPTQ runs symmetric with no activation reordering, asymmetric with no activation reordering, and asymmetric with group activation reordering; AWQ runs asymmetric with no activation reordering. DynQuant's quantizer is asymmetric and does not reorder columns, which is a property of the method rather than a recipe flag, so the panel records no scheme for its arms. Where a comparison above pairs a symmetric arm against an asymmetric one its delta spans two differences at once -- how the bits were allocated, and whether a zero point was stored per group -- so a large gap between those two arms is not on its own evidence about allocation. The comparison that isolates it is in this panel: gptq_3b and gptq_3b_asym_noao are the same method at the same byte anchor and differ in the scheme alone, so the difference between those two rows is the scheme and nothing else.
• Storage, measured; throughput, not. The number reported here is bytes on disk and execution match. This card makes no claim about decode speed or peak VRAM against an fp16 baseline, because this panel did not measure either.
• One task. Execution match on held-out text-to-SQL is what was scored. It says nothing about how this arm behaves on anything else, and a quantization that holds one task can lose another.

Install

This directory is packed, so transformers alone cannot open it -- it needs DynQuant's HfQuantizer, which the package registers. Prebuilt CUDA kernels come with it where a wheel exists for your platform, and it falls back to a pure-torch path where one does not.

pip install dynquant

Source, format spec, and the allocator that produced this arm's bit map: https://github.com/kambojvikram/dynquant

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

import dynquant

dynquant.register_hf_quantizer()

model = AutoModelForCausalLM.from_pretrained("VikramPal/mistral-7b-instruct-v0.3-DynQuant-4bit", device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained("VikramPal/mistral-7b-instruct-v0.3-DynQuant-4bit")

Provenance

• panel model: /workspace/runs/s4/mistral7b-v03.text2sql/merged
• parameters counted: 7,248,023,552
• byte target this arm was allocated against: 3,964,674,048 B
• fine-tune: 2472 steps, train loss 0.0540, 3.8 h
• fine-tune commit: 5959fe04b7db0512eaf7567d77a43f3df5860651
• evaluation: 2,454 problems in 12 min