mohashari’s gists

mohashari / snippet-1.py

Created July 5, 2026 01:00

Speculative Decoding in Production: Accelerating LLM Inference via Custom Draft Models — code snippets

	import torch

	def verify_draft_tokens(
	target_probs: torch.Tensor, # Shape: (K+1, vocab_size)
	draft_probs: torch.Tensor, # Shape: (K, vocab_size)
	draft_tokens: torch.Tensor, # Shape: (K,)
	) -> tuple[torch.Tensor, int]:
	"""
	Executes rejection sampling to verify K draft tokens against target probabilities.
	Returns a tensor of accepted tokens and the index of the first rejected token.

mohashari / snippet-1.hcl

Created July 4, 2026 01:03

Automating Database Schema Migrations in CI/CD: Multi-Tenant Zero-Downtime Rollouts with Atlas and GitHub Actions — code snippets

	env "local" {
	src = "file://schema.hcl"
	dev = "docker://postgres/15/dev?search_path=public"
	migration {
	dir = "file://migrations"
	}
	format {
	migrate {
	diff = "{{ sql . \" \" }}"
	}

mohashari / snippet-1.yaml

Created July 4, 2026 01:02

Fine-Tuning LoRA Adapters on Kubernetes: Orchestrating Distributed Training with Ray and Ludwig — code snippets

mohashari / snippet-1.go

Created July 4, 2026 01:02

Mitigating SSRF in Serverless: Designing a Dynamic Envoy Proxy Sidecar for AWS ECS Tasks — code snippets

mohashari / snippet-1.sql

Created July 4, 2026 01:01

Custom MySQL Binlog Parsers: Building a Real-Time Outbox Pattern Processor in Go — code snippets

	CREATE TABLE IF NOT EXISTS outbox_events (
	id BIGINT UNSIGNED AUTO_INCREMENT PRIMARY KEY,
	aggregate_type VARCHAR(255) NOT NULL,
	aggregate_id VARCHAR(255) NOT NULL,
	event_type VARCHAR(255) NOT NULL,
	payload JSON NOT NULL,
	created_at TIMESTAMP(6) DEFAULT CURRENT_TIMESTAMP(6),
	INDEX idx_created_at (created_at)
	) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin;

mohashari / snippet-1.go

Created July 4, 2026 01:00

Debugging HTTP/2 Flow Control Exhaustion: Analyzing gRPC Stream Stalls in Production — code snippets

	// Typical goroutine block stack trace found during a pprof capture:
	// goroutine 4821 [semacquire, 10 minutes]:
	// google.golang.org/grpc/internal/transport.(*controlBuffer).get(0xc0001a23c0, 0x1)
	// /go/pkg/mod/google.golang.org/grpc@v1.64.0/internal/transport/controlbuf.go:394 +0x95
	// google.golang.org/grpc/internal/transport.(*http2Server).Write(0xc00021c000, 0xc0001bc300, 0xc0003c2000, 0x0)
	// /go/pkg/mod/google.golang.org/grpc@v1.64.0/internal/transport/http2_server.go:882 +0x14c
	// google.golang.org/grpc.(*serverStream).SendMsg(0xc0003c2000, {0x18ba620, 0xc00045e4c0})
	// /go/pkg/mod/google.golang.org/grpc@v1.64.0/stream.go:1422 +0x125
	// main.(*ReporterServer).StreamReports(0xc0002ba140, 0xc00028e180, 0xc0003c2000)
	// /workspace/main.go:84 +0xbc

mohashari / snippet-1.hcl

Created July 3, 2026 12:54

Preventing Credential Leakage: Automated Just-In-Time AWS IAM Session Token Generation in GitHub Actions using OIDC and Vault — code snippets

	resource "vault_jwt_auth_backend" "github" {
	description = "JWT authentication backend for GitHub Actions"
	path = "github-actions"
	oidc_discovery_url = "https://token.actions.githubusercontent.com"
	bound_issuer = "https://token.actions.githubusercontent.com"
	}

mohashari / snippet-1.txt

Created July 3, 2026 12:53

Building Custom TensorRT-LLM Plugins for Low-Latency FP8 Inference at Scale — code snippets

	#pragma once
	#include <NvInferRuntime.h>
	#include <vector>
	#include <string>

	namespace tensorrt_llm::plugins {

	class FusedSwiGLUQuantPlugin : public nvinfer1::IPluginV2DynamicExt {
	public:
	FusedSwiGLUQuantPlugin(float scale, bool has_dynamic_scale);

mohashari / snippet-1.txt

Created July 1, 2026 01:03

Designing an Anomaly Detection Engine for High-Cardinality Metrics Using Holt-Winters in Rust — code snippets

	// The core math engine for additive Holt-Winters triple exponential smoothing.
	// Maintains rolling level, trend, and seasonal components in a memory-efficient manner.

	pub struct HoltWinters {
	alpha: f64, // Level smoothing parameter (0 < alpha < 1)
	beta: f64, // Trend smoothing parameter (0 < beta < 1)
	gamma: f64, // Seasonal smoothing parameter (0 < gamma < 1)
	period: usize, // Seasonality period length (e.g. 288 steps for 24h at 5m resolution)
	level: f64, // Current estimate of the base level
	trend: f64, // Current estimate of the trend direction

mohashari / snippet-1.py

Created July 1, 2026 01:02

Optimizing Key-Value Cache Management in LLM Serving with PagedAttention and CUDA IPC — code snippets

	import torch
	from typing import List, Set, Dict

	class PhysicalBlockAllocator:
	def __init__(
	self,
	num_blocks: int,
	block_size: int,
	num_layers: int,
	num_heads: int,

MOH ASHARI MUKLIS mohashari