qwerky-vllm-models 0.2.23

vLLM plugin for Qwerky AI MambaInLlama hybrid models

Qwerky vLLM Models

A vLLM plugin for serving Qwerky AI's MambaInLlama hybrid models without the --trust-remote-code flag.

Installation

pip install vllm qwerky-vllm-models

Usage

After installing, serve Qwerky models with vLLM:

vllm serve QwerkyAI/Qwerky-Llama3.2-Mamba-3B-Llama3.3-70B-base-distill --max-model-len 4096

The plugin automatically registers the model architecture with vLLM on import.

Supported Models

• QwerkyAI/Qwerky-Llama3.2-Mamba-3B-Llama3.3-70B-base-distill

How It Works

This package uses vLLM's plugin system (vllm.general_plugins entry point) to register the MambaInLlama model architecture. This means:

• No fork of vLLM required
• No --trust-remote-code flag needed
• Works with standard vLLM installation
• Uses vLLM's native Triton-accelerated Mamba kernels

Requirements

• Python >= 3.10
• vLLM >= 0.14.0
• PyTorch >= 2.0.0

Changelog

0.2.23

• CRITICAL FIX: Wrong in_proj split order causing gibberish output
• Reference implementation uses: [z(d_inner), x(d_xb), B(d_xb), C(d_inner), dt(dt_rank)]
• Our code incorrectly had: [z(d_inner), x(d_inner), B(d_xb), C(d_xb), dt(dt_rank)]
• x is d_xb (needs repeat_kv expansion), C is d_inner (already full size)
• Fixed _prefill and _decode_step to handle x/C dimensions correctly

0.2.22

• FIX: Double bias in Mamba dt computation (partial fix, not root cause)
• Removed redundant bias addition in _ssm_scan and _ssm_state_update

0.2.21

• FIX: Dtype mismatch in rotary position embeddings
• Cast cos/sin to match q's dtype before applying rotation
• Fixes RuntimeError: expected scalar type Float but found BFloat16 in Q×K matmul

0.2.20

• FIX: Dtype mismatch in attention matmul
• After softmax (computed in float32), convert to v.dtype instead of q.dtype
• Fixes RuntimeError: expected scalar type Float but found BFloat16

0.2.19

• FIX: Handle vLLM warmup where seq_len exceeds KV cache size
• During warmup/autotune, max_num_batched_tokens=8192 but cache only holds 2048
• Skip KV caching when tokens don't fit, allowing warmup to complete

0.2.18

• Added extensive debug logging to diagnose attention layer shape issue
• Logs: input shape, batch_size, seq_len, Q/K/V shapes, rotary output, KV cache shapes

0.2.17

• Added debug logging in MHADecoderLayer to trace tensor shapes

0.2.16

• Fixed attention layer to handle vLLM's flattened 2D tensor format
• vLLM passes [total_tokens, hidden] but attention needs [batch, seq, hidden]
• Added automatic batch dimension handling in MHADecoderLayer

0.2.15

• Fixed attention layer KV cache shape mismatch
• Removed incorrect tensor transpositions in KV cache assignment

0.2.14

• Fixed mamba_config.json loading - removed local_files_only=True restriction
• Now properly downloads mamba_config.json from HuggingFace Hub if not cached
• Added more detailed logging for config loading

0.2.13

• CRITICAL FIX: Load mamba_config.json for attn_layers, d_inner, d_xb
• MambaInLlama models store Mamba-specific config in separate mamba_config.json file
• Main config.json has model_type: "llama" without Mamba params
• Fixed: Model was treating ALL layers as Mamba (attn_layers=[]) because config wasn't loaded
• Added better logging for weight loading diagnostics
• Attention layers at indices [3, 8, 13, 18, 23, 27] now properly recognized

0.2.12

• CRITICAL FIX: Corrected d_xb default to match qwerky-distill PR #81
• d_xb = num_key_value_heads * head_dim (GQA-style, e.g., 8×128=1024 for 8B)
• Fixed in_proj split: [z(d_inner), x(d_inner), B(d_xb), C(d_xb), dt(dt_rank)]
• Added repeat_kv expansion for C (same as B) in Mamba1 architecture
• Fixed head count: num_heads = d_inner // d_state after B/C expansion

0.2.11

• CRITICAL FIX: Changed d_inner default from intermediate_size to hidden_size
• MambaInLlama Mamba layers use d_inner = hidden_size, not intermediate_size
• Fixed d_xb default: hidden_size // 16 (was hidden_size // 4)
• This fixes the shape mismatch for all Mamba layer weights (A_log, D, conv1d, dt_proj, in_proj, out_proj)

0.2.10

• Added debug logging to weight loading to diagnose parameter mapping issues
• Logs first 20 model params, first 20 checkpoint weights, and all skipped weights

0.2.9

• Fixed weight loading: split fused mha.in_proj into separate q/k/v projections
• Renamed mha.out_proj to o_proj for checkpoint compatibility
• Should now load all ~395 parameters instead of just 163

0.2.8

• Fixed dtype mismatch in SSM scan: F.softplus/torch.exp compute in float32, now cast back to original dtype
• This caused "expected BFloat16 but found Float" error in einsum

0.2.7

• Fixed tensor broadcasting bug in _ssm_scan: A.unsqueeze(0).unsqueeze(-1) -> A.unsqueeze(0).unsqueeze(2)
• This caused shape mismatch (8192 vs 16) during SSM discretization

0.2.6

• Added embed_input_ids method required by vLLM's VllmModelForTextGeneration interface
• This was the root cause of "This model does not support --runner generate" error

0.2.5

• Fixed vLLM runner detection: added MambaInLlamaMambaForCausalLM alias for HF config compatibility
• Added proper protocol inheritance (HasInnerState, IsHybrid) from vllm.model_executor.models.interfaces
• Fixed class variable type hints (ClassVar[Literal[True]]) for vLLM model inspection
• Simplified model registration code

0.2.4

• Complete architecture rewrite with explicit state cache management
• Separate prefill and decode paths for Mamba layers
• Grouped-head Mamba support (num_xb_head, num_C_head, repeat_group)
• Pure PyTorch SSM implementation (preparing for vLLM Triton op integration)

0.2.3

• Fixed d_xb default value computation in configuration
• Removed unsupported device/dtype kwargs from RMSNorm calls

0.2.2

• Fixed vLLM 0.14+ compatibility issues with Mamba ops API

0.2.1

• Updated README, removed SFT model reference

0.2.0

• Initial public release with vLLM plugin system integration

License

Apache 2.0