JyotinderSingh

"""GPTQ calibration wall-time benchmark: batched vs unbatched forwards.

Builds a synthetic 4-block transformer (d_model=512, ffn=2048, EinsumDense
attention projections) and runs full GPTQ calibration over ~128 synthetic
calibration sequences of length 128 (no downloads).

"old" == calibration_batch_size=1 (reproduces the pre-change per-sample
forward pattern). "new" == calibration_batch_size=8 (the new default). The
Cholesky inverse-Hessian change is orthogonal and per-layer (negligible vs the
forward passes), so this isolates the calibration-forward speedup from batching.

JyotinderSingh

===== A. STRING quantize('int4') =====
dtype_policy: int4/128_from_float32
_int4_block_size: 128
  var bias             shape=(64,) dtype=float32
  var kernel           shape=(256, 32) dtype=int8
  var kernel_scale     shape=(2, 64) dtype=float32
  var kernel_zero      shape=(2, 64) dtype=int8
  var g_idx            shape=(256,) dtype=float32
output[0,:5]: [-0.75887 -0.52892 -0.41521 -1.82848 -0.12788]
output sum: -15.898343086242676

JyotinderSingh

"""Evidence: OLD vs NEW Keras AWQ fidelity.

Compares the previous Keras AWQ algorithm (running-MAX activation statistic,
scale = x_max**ratio only, no weight clipping) against the reference-aligned
implementation on this branch (running-MEAN activation statistic, reference
scale formula x_stat**ratio / w_stat**(1-ratio), plus AutoAWQ-style per-group
weight clipping).

Two metrics:
  (a) Per-layer activation-weighted reconstruction error

JyotinderSingh

#!/usr/bin/env python3
"""One-off benchmark for Data.content_hash().

Creates temporary datasets on disk, then times content_hash().

Usage:
    python bench_content_hash.py
    python bench_content_hash.py --files 50000 --total-mb 500
"""

JyotinderSingh

class Dense(Layer):
    
    ...
    
    def get_gptq_handler(self) -> "GPTQHandler":
        """Provides a GPTQHandler for a standard Dense layer."""
        return GPTQHandler(
            kernel=self.kernel,
            bias=self.bias,
            rows=self.kernel.shape[0],

JyotinderSingh

package main

import (
	"context"
	"encoding/json"
	"fmt"
	"log"
	"net/http"

	"github.com/dicedb/dicedb-go"

JyotinderSingh

package main

import (
	"context"
	"encoding/json"
	"fmt"
	"log"
	"math/rand"
	"net/http"
	"os"

JyotinderSingh

conda activate <env_name>

conda install -c conda-forge jupyter_contrib_nbextensions

jupyter nbextensions_configurator enable --user

jupyter contrib nbextension install --user

jupyter nbextension enable hinterland/hinterland

JyotinderSingh

Cypress.Commands.add('loginOkta', () => {
  const optionsSessionToken = {
    method: 'POST',
    url: Cypress.env('session_token_url'),
    body: {
      username: Cypress.env('username'),
      password: Cypress.env('password'),
      options: {
        warnBeforePasswordExpired: 'true'
      }

JyotinderSingh

diff --git a/compiler/build.gradle b/compiler/build.gradle
index 60d3a436f..409df7a03 100644
--- a/compiler/build.gradle
+++ b/compiler/build.gradle
@@ -31,6 +31,7 @@ def addLibraryIfNotLinked = { libName, argList ->
 }
 
 def String arch = rootProject.hasProperty('targetArch') ? rootProject.targetArch : osdetector.arch
+// arch = arch.replace('osx_arm-v8', 'osx_aarch64')
 def boolean vcDisable = rootProject.hasProperty('vcDisable') ? rootProject.vcDisable : false