[Quantization] Enable humming wNaM asymmetric quant (zero_point) with compressed-tensors (#46528)

This commit is contained in:
HDCharles
2026-09-23 10:23:11 -04:00
committed by GitHub
parent 88afb9dcac
commit 9f07d023d0
12 changed files with 164 additions and 34 deletions
@@ -1,5 +1,5 @@
model_name: "nm-testing/Qwen3-4B-mixed-quant-RTN-wNaM"
accuracy_threshold: 0.78
model_name: "nm-testing/Qwen3-4B-mixed-quant-RTN-wNaM-a4"
accuracy_threshold: 0.70
num_questions: 1319
num_fewshot: 5
server_args: >-
@@ -0,0 +1,8 @@
model_name: "nm-testing/Qwen3-4B-mixed-quant-RTN-wNaM-asym"
accuracy_threshold: 0.85
num_questions: 1319
num_fewshot: 5
server_args: >-
--enforce-eager
--max-model-len 4096
--linear-backend humming
@@ -1,3 +1,4 @@
gpt-oss-20b-humming.yaml
Qwen3.5-35B-A3B-experts-int8-humming-act-int8.yaml
Qwen3.5-35B-A3B-int6wc-hadamard-humming-act-int8.yaml
Qwen3-4B-mixed-quant-RTN-no-w4a4-humming.yaml
@@ -4,4 +4,4 @@ Qwen3-30B-A3B-GPTQ-Int4-humming-act-int8.yaml
Qwen3-30B-A3B-AWQ-humming.yaml
Qwen3.5-35B-A3B-experts-int8-humming-act-int8.yaml
Qwen3.5-35B-A3B-int6wc-hadamard-humming-act-int8.yaml
Qwen3-4B-mixed-quant-RTN-humming.yaml
Qwen3-4B-mixed-quant-RTN-no-w4a4-humming.yaml